Mixture-Of-Expert Language Models

Mixture-of-Experts (MoE) language models suffer from large parameter counts, which create a significant memory bottleneck. Expert pruning is the most direct approach for reducing this parameter count, yet existing methods make pruning decisions for each expert independently, and assume experts' contributions are purely additive. In reality, expert usage in MoEs is inherently cooperative. We derive HOPE (Higher-Order Pruning of Experts), a second-order pruning objective which provably minimizes an upper bound on the error resulting from pruning. We show that REAP (a state-of-the-art first-order pruning method) is a special case of HOPE where interaction terms are ignored. Across three frontier MoE models (up to 122B parameters), two distinct calibration sets, and multiple benchmarks (including math, instruction following, coding, and an agentic suite), we demonstrate that HOPE produces better pruning decisions than existing methods, and its advantage is most pronounced at high pruning rates and on challenging agentic workloads. At 50% pruning, HOPE outperforms all baselines and achieves an average rank of 1.58 out of 5 methods (versus 2.42 for the next-best method, REAP), with gains of up to +6.1% on agentic coding. Over all conditions, HOPE again achieves the best average rank and surpasses every other method in the majority of head-to-head comparisons. By preserving cooperative expert structure that first-order methods ignore, HOPE enables aggressive compression with minimal degradation, particularly on complex tasks where diverse expert combinations are invoked over long sequences.

Scaling large language models efficiently has motivated sparse capacity mechanisms such as Mixture-of-Experts and, more recently, conditional memory: token-indexed embedding tables that augment the backbone with cheap parametric lookups. Existing memory-embedding methods retrieve via a deterministic function of the surface form, which collapses different contextual senses of the same token (e.g., python the language vs. the animal) into a single fixed entry. We introduce Mixture of Memory Embeddings (MoME), a context-aware memory mechanism that replaces each token's single memory row with a mixture of M slots and uses a learned gate over the hidden state to choose which slots to read at each position. In controlled pretraining experiments across nanochat, Llama-3/MobileLLM, and Qwen3 backbones, MoME improves over Value Embedding, Bigram, and STEM baselines in iso-parameter and iso-training-FLOP settings, shows a more promising memory-size scaling trend at sub-billion scale, and remains efficient in training and inference. Qualitative routing analyses on polysemous tokens further suggest that the learned mixture exhibits a degree of semantic interpretability, dispatching the same surface token to distinct memory slots under different senses.

In this work, we introduce Instella-MoE, a fully open Mixture-of-Experts (MoE) language model with 16 billion total parameters and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. Instella-MoE combines a sparsely activated MoE design with architectural and system-level innovations, including Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity, enabling efficient large-scale training and inference. The model is developed through a multi-stage pipeline comprising pre-training, mid-training, long-context extension, supervised fine-tuning with feedback-driven data curation, direct preference optimization, and reinforcement learning with Multi-Teacher On-Policy Distillation. Instella-MoE achieves an average score of 76.7 across standard pre-training benchmarks, outperforming prior fully open models including OLMo-3-7B, SmolLM3-3B, and OLMoE-1B-7B, while remaining competitive with open-weight MoE and dense baselines at comparable active-parameter scales, including Moonlight-16B-A3B and Qwen3.5-4B. After post-training, our final Think checkpoint achieves an average score of 73.2 across instruction-following, reasoning, math, coding, and chat benchmarks, outperforming both fully open and open-weight models with comparable or larger active parameter counts in our evaluation. To support transparent and reproducible research, we release the complete Instella-MoE model flow, including model weights, training configurations, data mixtures, and training code. Together, these contributions establish Instella-MoE a strong, fully open foundation for efficient, high-performing MoE models and reproducible research.

Mixture-of-experts (MoE) architectures scale large language models efficiently, but they demand massive GPU memory. To cope with such demand, models are commonly compressed to reduce their memory footprint. Residual sparsification is a representative compression technique that decomposes each projection matrix of an expert into a shared base matrix and per-expert residual matrix, and then compresses the residuals. Existing sparsification methods compress each residual matrix independently by minimizing its compression error, thereby minimizing the error of each projection matrix. However, our analysis shows that this objective is misaligned with preserving model accuracy after compression. In an expert, the final output is produced through computations coupled across multiple projections and hidden representations. Therefore, even small errors in individual matrices can propagate through hidden representations and projection interactions, leading to large expert output errors and accuracy degradation. To address this misalignment, we propose PARSER, a new residual sparsification method that shifts the compression objective from minimizing isolated matrix errors to preserving the expert output error. PARSER achieves this by introducing output importance, which measures the actual contribution to the expert output error. Our experiments show that, compared with existing methods, PARSER narrows the accuracy gap to the uncompressed model by 1.41$\times$ on Qwen and 1.44$\times$ on DeepSeek, while achieving the same peak memory reduction. Our code is available at https://github.com/OSSS-KU/PARSER.

Mixed-precision quantization (MPQ) assigns a different bitwidth to each linear layer of a large language model (LLM) to minimize the quantization-induced quality loss under a fixed budget, but Mixture-of-Experts (MoE) models contain these layers in every expert of every MoE block, so the allocation space grows far larger than in a dense model. Existing methods either allocate within each block under a uniform per-block budget, or allocate across blocks through an additive proxy, and neither directly optimizes a model-level objective over the choices that couple the blocks. We propose Q-Strata, a bi-level allocator that ranks within-block assignments with a cheap proxy and allocates across blocks with a model-level objective evaluated on the assembled quantized model. Its inner stage caches a Pareto frontier of candidates per block over finely spaced budgets, leaving the outer stage to set one budget per block instead of a bitwidth for every linear layer. With the search reduced to one budget per block, the outer stage optimizes this model-level objective directly, capturing the inter-block coupling that additive proxies miss. On Mixtral-8x7B-Instruct, Qwen1.5-MoE-A2.7B, and DeepSeek-V2-Lite, Q-Strata consistently achieves lower WikiText2 perplexity than uniform-bitwidth GPTQ and the state-of-the-art MoE MPQ methods MxMoE and GEMQ in the low-bit regime. The code is available at https://github.com/snu-mllab/Q-Strata/tree/main.

We introduce A.X K2, a 688B-parameter Mixture-of-Experts (MoE) language model trained from scratch as a high-performance foundation for \emph{agentic} applications. Trained on approximately 8.5T tokens---fewer than its predecessor, A.X K1---on a smaller but higher-quality mixture with substantially expanded agentic and software-engineering data, it nonetheless improves over A.X K1 across the board, by over 30 percentage points on some benchmarks, reflecting large gains in token efficiency. To support long contexts efficiently, we introduce Sparse Gated Attention (SGA), which combines sparse attention with gated attention, and adopt Gated Norm (GN) to stabilize large-scale training. SGA is trained natively at 128K through a \emph{sparse} indexer warmup that optimizes the indexer against its own sparse top-$k$ selection rather than the dense attention distribution, making adaptation markedly cheaper: each query reads only 2,048 positions, yet long-context quality is unchanged and A.X K2 scores 94.6 on RULER out to 256K. The outlier suppression of GN in turn keeps 4-bit NVFP4 serving within one point of FP8 accuracy. A simple yet effective Think-Fusion recipe further lets users switch between thinking and non-thinking modes within a single unified model. Extensive evaluations show that A.X K2 performs competitively against strong open-weight baselines, matching or exceeding them on math and Korean-language benchmarks.

Large language models (LLMs) have shown strong potential for financial analysis and trading, but direct trading remains challenging because the predictive capabilities required can vary across assets, decision fields, and market conditions. Existing LLM-based trading systems either coordinate human-defined external experts or adopt conventional internal Mixture-of-Experts (MoE) routers that do not directly evaluate how individual experts contribute to trading decisions. Moreover, these routers receive no direct signal indicating when an inactive expert has become more suitable as market conditions change. We find that native router scores poorly reflect how much individual experts improve trading decisions, frequently leaving better alternatives unselected. We further reveal that token-specific expert usefulness exhibits a compact low-dimensional structure. Based on these findings, we propose TradingMoE, a trading-oriented sparse MoE that augments a frozen dense LLM with lightweight residual experts. We introduce a Query-Key router that represents the expertise required by each token under the current market context as a low-dimensional query and matches it with learnable expert keys. We further propose a sparse expert selection update mechanism that samples a few inactive experts during training and estimates whether they should replace the weakest expert in the current Top-k route. This mechanism enables the router to update expert selection as market conditions change while preserving sparse computation. Experiments against 22 baselines on stock and cryptocurrency markets show that TradingMoE improves cumulative return over the best-performing baselines by 30.89% and 30.7%, respectively. Rolling paper-trading experiments further demonstrate that its advantage persists under forward-only deployment.

We introduce Motif 3, a decoder-only Mixture-of-Experts language model with 314 billion total parameters and 13.2 billion activated per token. Each sparse MoE layer contains 384 routed experts, with eight selected per token. This fine-grained sparsity provides substantial expert capacity while limiting computation. Motif 3 is built around Grouped Differential Latent Attention (GDLA), which integrates grouped differential attention with the compressed key-value representation of Multi-head Latent Attention. The architecture further incorporates modified manifold-constrained hyper-connections, Expert Specific PolyNorm activations, and multi-token prediction to improve optimization stability, expert specialization, and inference efficiency. We pretrain Motif 3 on approximately 12.5 trillion tokens spanning web documents, STEM, code, mathematics, multilingual content, and domain-specialized corpora. Expert-balancing and numerical-stabilization techniques support stable training at scale, while selective MXFP8 computation and communication, memory-efficient fused kernels, and window-aware context parallelism enable training with context lengths up to 256K tokens. Our post-training pipeline combines general supervised fine-tuning, six specialist teachers trained with reinforcement learning, a software-engineering teacher trained with supervised fine-tuning, and Multi-teacher On-Policy Distillation. The resulting unified model consolidates complementary capabilities in reasoning, coding, tool use, professional work, long-context understanding, calibrated abstention, and instruction following. Across a broad evaluation suite, Motif 3 demonstrates competitive performance against leading open weight models, including strong results on long-horizon agentic tasks, mathematical reasoning, scientific knowledge, and hallucination-sensitive evaluation.

Safety alignment in multilingual models is uneven: a model that reliably refuses a harmful request in English will often comply with the same request in a lower-resource language. We trace this gap mechanistically in sarvam, an Indic-multilingual mixture-of-experts reasoning model, and find it is not a failure to detect harm. Harm is encoded as an internal direction that is nearly language-invariant in mid-network (English-vs-Indic cosine ${\approx}0.9$ at $L11$), and steering that direction upstream causally controls refusal. But the detection direction is orthogonal to the change that actually writes the refusal, which is late and assembled over the course of generation rather than read off in a single forward pass. We attribute the write to a specific, localizable circuit, a mixture-of-experts writer held in check by an attention opposer and price every way of intervening on it: damping the opposer is cheap and effective, amplifying the writer is a cost wall, and surgical edits to the responsible heads do nothing. The circuit's organization, and the gradient method that exposes it, recur in a second, unrelated MoE model, while the lever's strength is architecture-specific. The result is a cost-measured map of where a multilingual safety repair can land, and what it costs

Mixture-of-Experts (MoE) language models deliver high capacity at low per-token compute, but deploying them cheaply requires compressing their many expert weight matrices. Expert pruning (e.g., REAP) and merging reduce cost but sacrifice accuracy and require retraining the router; low-rank delta decomposition of experts (e.g., D^2-MoE) preserves all experts and the router, but degrades sharply as the expert count grows because a single shared component cannot approximate many near-orthogonal experts. Because MoE expert weights are near-orthogonal, a single shared component (as in prior delta decomposition) scales poorly with the expert count; we show that experts nonetheless organize into functional co-activation communities that are decoupled from weight similarity. Building on this, we introduce LorExperts, a router-preserving compression method that clusters experts, keeps one full-precision dominant per cluster, and represents the remaining members as low-rank corrections to their local dominant. LorExperts retains all experts and the original router (no router retraining). At ~50% expert compression on Qwen3-30B-A3B and Gemma-4-26B-A4B, LorExperts preserves downstream accuracy and perplexity better than the baselines on most of the tasks; the margin over D^2-MoE grows with expert count E. We further give a reconstruction fine-tuning procedure for LorExperts, and BTExperts, a tree organization of dominants and corrections that enables inference-time amortization of shared computation.

Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learning rate decays more rapidly with compute. For model--data allocation, IsoFLOP analysis reveals a slight data-side tilt: the optimal token budget grows faster than activated model-side computation. For MoE architecture, larger scales increasingly favor larger expert pools at fixed activated capacity, while moderate expert granularity remains consistently effective and the preferred fraction of activated capacity assigned to shared experts remains stable across scales. Guided by these findings, we train LLaDA MoE v2, a 30B-A3B dLLM, from scratch on 23.5T tokens. With approximately 65% as many pretraining tokens as Qwen3, LLaDA MoE v2 approaches Qwen3 on several knowledge, reasoning, and coding benchmarks. After supervised fine-tuning alone, it outperforms SDAR Chat on seven of eight reasoning and coding benchmarks and remains close to Qwen3 on several tasks. These results establish practical scaling laws and design principles for MoE dLLMs.

Parameter-efficient fine-tuning (PEFT) enables efficient adaptation of large language models, but existing MoE-based PEFT methods typically improve capacity by storing multiple full LoRA experts, causing adapter storage to grow linearly with the number of experts and restricting adaptation to a fixed expert pool. We ask whether MoE-based PEFT can produce instance-specific adaptations without explicitly storing a separate LoRA module for each expert. To address this gap, we propose MoEGen, an adaptation framework that shifts MoE-based PEFT from expert selection to expert-conditioned parameter generation. Instead of storing each expert as a full LoRA adapter, MoEGen represents each expert as a small learnable vector, termed an expert code. It routes each input over these vectors and uses their weighted combination to condition a lightweight hypernetwork that generates input-specific low-rank updates. This design decouples expert capacity from adapter storage while enabling instance-conditioned adaptation. Experiments on eight commonsense reasoning benchmarks show consistent improvements over strong static and MoE-based PEFT baselines across three backbones. MoEGen also performs strongly in joint medical and legal-domain adaptation.

Native computer use offers a general interface for agents to operate almost any software available to people, but requires long-horizon state tracking, large-scale interactive experience, and learning from sparse yet verifiable outcomes. We introduce Qwen-CUA, a native computer-use agent with a 397B-A17B Qwen mixture-of-experts backbone. It observes only screenshots and acts through keyboard and mouse events, without DOM trees, accessibility metadata, or task-specific APIs. Its scaffold maintains up to 20 active screenshots and folds older visual history in fixed-size blocks to retain recent evidence while preserving reusable prompt prefixes. For training, we build a cloud rollout fleet with access to nearly 100,000 vCPUs and tens of thousands of concurrent environments, construct approximately 40,000 verifiable tasks, and collect personalized long-horizon workflows across everyday and professional software. We optimize complete trajectories with verifiable rewards and trajectory slicing, while iterative training runs refresh supervised data and recalibrate reinforcement-learning tasks. Across eight benchmarks, Qwen-CUA outperforms Qwen3.7 and remains competitive with leading proprietary systems, reaching 86.2 on OSWorld-Verified and 18.5/48.4 binary/partial completion on OSWorld 2.0. Scaling the same recipe to a model with over one trillion parameters yields Qwen-CUA-Max, improving these scores to 87.6 and 21.2/53.3. Qwen-CUA also reduces RedTeamCUA attack success from 36.6 to 16.4 relative to Qwen3.7. Efficiency analyses, a browser deployment, and Bash-augmented experiments further characterize practical behavior. These results establish native computer use as a broadly capable agent foundation and highlight scalable verifiable interaction and hybrid tool use as key directions.

Modern large language models scale successfully by pairing capacity growth with efficiency, keeping per-token and deployment costs under control as capacity grows. AIGC Foundation Models (AFMs), especially diffusion-transformer backbones, have begun to adopt sparse experts, but recent efforts mostly enlarge total parameter counts and sparsity ratios without importing the efficiency mechanisms that made LLM scaling practical, so generation quality is seldom balanced against training and deployment cost. This raises a natural question: can the architectural principles behind efficient LLM scaling be adapted to AFMs in a more balanced way? We introduce ModernMOE (MMOE), a modernization of SiT-style diffusion transformers that systematically adapts routed experts, shared and lightweight experts, gate-residual routing, and attention-residual information reuse to AIGC generation. Rather than treating MoE as a single plug-in replacement, MMOE studies how different modern expert components affect convergence, efficiency, and generation quality when composed inside a diffusion transformer. Every experiment in this paper is trained on a single eight-GPU H100 node with batch size 256 for 400k steps, an accessible single-machine budget. Under matched training and sampling protocols and at this budget, MMOE reaches lower FID at every recorded checkpoint, that is, it converges faster per training step, than dense and intermediate sparse-expert baselines, and among the sparse variants it attains the best quality-cost balance. Routing analysis further shows stable expert specialization across depth, substantial use of lightweight routes, and modest step-to-step routing changes during denoising. These results suggest that AFMs can follow the balanced scaling path of LLMs by importing proven efficiency designs, rather than by simply increasing total parameters and sparsity ratios.

Massively multilingual automatic speech recognition (ASR) models covering hundreds of languages must maintain robust performance across diverse linguistic and acoustic conditions. However, these models often encounter the curse of multilinguality, where model capacity is diluted across languages. To address this challenge, we propose Mixture of Language Group Experts (MoLGE), built upon speech self-supervised models (S3Ms). MoLGE assigns dedicated expert modules to clusters of similar languages, reducing the number of required submodules compared to conventional language-specific Mixture-of-Experts (MoE) schemes. It further integrates a hierarchical Low-Rank Adaptation (LoRA) strategy into the disentangled acoustic and linguistic components of the S3M architecture, enabling efficient modeling of language-specific characteristics while maintaining parameter efficiency. Further, we investigate the impact of language grouping strategies based on both linguistic and data-driven criteria on overall performance, providing an interpretable perspective on how language structure influences scalability in multilingual speech systems. In experiments, we evaluate MoLGE on a multilingual benchmark encompassing 495 languages. Results demonstrate that MoLGE consistently outperforms dense multilingual baselines with a minimal increase in trainable parameters. Notably, these language grouping strategies yield substantial improvements for both phonetic and orthographic aspects of ASR modeling. Our findings suggest that structured language specialization provides an effective pathway for massively scaling language coverage of multilingual ASR.

Mixture-of-Experts (MoE) architectures have been widely adopted in large language models, yet parameter-efficient fine-tuning (PEFT) for MoE models remains underexplored. Existing PEFT methods for MoE either ignore router priors with uniform adapters, reducing efficiency and risking forgetting, or rely on static expert selection, limiting per-token capacity and cross-expert feature learning. In this paper, we make the first attempt to fine-tune MoE models with MoE-style low-rank adaptation: our method, entitled MoE$^2$-LoRA, deeply couples the pretrained expert specialization with task-specific adaptivity via a dual-channel Routing-Conditioned Projection (RCP) module, which reuses base router activations to inform LoRA routing. We further introduce a single global LoRA expert pool shared across all layers, enabling model-wide adaptation with emergent layer-wise affinities and balanced expert utilization. MoE$^2$-LoRA simultaneously benefits from the advantages of prior reuse, dynamic adapter routing, and model-wide knowledge sharing. Evaluated on multiple MoE backbones with varying scales and expert granularities, MoE$^2$-LoRA consistently achieves state-of-the-art downstream accuracy while retaining stronger general capabilities.

We present Solar Open 2, a 250B-A15B Mixture-of-Experts language model built for long-horizon agentic tasks, scaled up from Solar Open 1 (Solar Open 100B). To hold entire agent trajectories in a single context, Solar Open 2 reaches a 1M-token window through a hybrid attention stack that interleaves one softmax layer among every three linear-attention layers, using no positional encoding and a gated delta rule extended to negative eigenvalues. To train at this scale under a fixed compute budget, we make training efficient in two ways: a stronger starting point, and higher-value data. For the starting point, we initialize Solar Open 2 from Solar Open 1, transferring the 5.69B-parameter shared skeleton that survives the architectural change and learning everything else through full pre-training. For the data, we curate for value per token: quality- and rarity-aware data curation and mixture-ratio optimization refine a 20T pool into a 10T mixture that, at equal token budget, outperforms the Solar Open 1 recipe. To build its agent skills, we train twelve domain specialists across purpose-built scenarios, then consolidate them into a single model by Multi-teacher On-Policy Distillation (MOPD). Against comparably sized open-weight models on English benchmarks, Solar Open 2 leads on MMLU-Pro, LiveCodeBench, and the APEX-Agents agentic suite, and stays competitive with the strongest (DeepSeek-V4-Flash and MiMo-V2.5) elsewhere. On Korean benchmarks, Solar Open 2 records the highest average of any model compared, including fast-tier closed APIs, and on Ko-GDPval, an in-house Korean officework-agent benchmark, it is competitive with DeepSeek-V4-Pro (1.6T) at less than a sixth of its size.

The strongest open-weight coding models are mixture-of-experts (MoE) networks: most of their size comes from large pools of "expert" subnetworks, of which only a few act on any token. That pool is why these models do not fit on the machines most developers own, yet for a user who only wants coding help, most experts encode abilities that will never be invoked. We ask how many experts can be removed, and which, by pruning two recent open-weight MoE models from different families (Qwen3.6-35B-A3B and Gemma-4-26B-A4B) under five selection strategies, judged the way a user would: by whether the model still writes correct code. Half the experts can be removed from either model with no statistically detectable loss on the primary code benchmark, and the damage lands almost entirely on abilities outside coding, the intended trade. But the winning strategy flips between the two models, so a recipe validated on one family cannot be assumed to work on another. We further show that perplexity, the metric much of the pruning literature leans on, can rate a broken model above an intact one; that a lightweight fine-tune recovers about half of what aggressive pruning loses; and that against quantizing the full model to the same memory, pruning wins only where quantization would have to drop below 3 bits per weight. Five attempts to overturn that crossover, with failure criteria fixed in advance (better calibration, guarded selection, causal expert importance, failure attribution, and an agentic evaluation letting each model repair its failures from execution feedback), all leave it standing; the last shows single-shot benchmarks overstate compression penalties broadly, as one repair turn erases the 2-bit quantization penalty entirely. Expert pruning works, but it demands per-model validation on the task the model will actually serve.

Mixture-of-Experts (MoE) is a popular class of large language models (LLMs), offering high efficiency and accuracy. However, in KV-cache-intensive serving scenarios, MoEs often exhibit a tension between the GPU memory requirements of the model weights and the growing KV cache. We propose PagedWeight, a novel management method for MoE LLM serving that dynamically quantizes MoE model's weights at runtime and balances expert-weight precision with the KV cache sizes. PagedWeight exposes and effectively navigates the complex tradeoff between the model's task accuracy, memory consumption, and throughput/latency. Across several memory-sensitive MoE serving scenarios, PagedWeight improves the quality-memory tradeoff over several existing quantization baselines. PagedWeight achieves FP16-equivalent accuracy with up to 72.0% GPU memory savings and 1.94$\times$ throughput improvement, and improves quality over quantization methods by up to 39.3% at a similar memory budget with at most 4.1% throughput loss.

Mixture-of-Experts (MoE) models scale capacity without proportional compute cost and have become a key architecture for frontier large language models (LLMs). Yet domain-specific post-training inherits an expert pool shaped by mixed-domain pre-training: a substantial subset of experts contributes little on the target domain, and standard supervised fine-tuning (SFT) leaves the composition of this pool unchanged. We propose a simple, budget-preserving pipeline that realigns the expert pool to the target domain before fine-tuning. Given a target domain, we (1) prune the experts with lowest domain-aligned saliency, (2) regrow the expert pool to its original size through perturbation-based expert expansion, and (3) apply standard SFT. The resulting model preserves the original expert count, parameter count, and inference cost. With a single frozen recipe and no per-domain hyperparameter tuning, UMoE consistently improves over direct sft across two MoE architectures (Qwen3-30B-A3B and Qwen3.5-35B-A3B), five domains (math, code, science, tool-use, and agentic coding), and 12 benchmarks. Representative improvements are 3.4 points in math average accuracy, 6.0 points on SWE-bench Verified. On a strong in-house math corpus, direct sft already surpasses Qwen3-30B-A3B-Thinking (82.81 vs.\ 81.06), yet UMoE further raises the average to 84.17, an additional 1.36 points, demonstrating robustness to a substantially stronger SFT regime. Data-scaling experiments further show that the gain persists as training data grows. Analysis reveals that the direct-SFT model allocates substantial routed-expert compute to a low-saliency subset that can be removed post hoc with little average degradation; UMoE turns this redundant capacity into useful domain capacity and achieves lower training loss, with gains spanning all difficulty levels in downstream evaluation.

We present Soofi S 30B-A3B, a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba Transformer foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense models for long-context, high-concurrency deployment. Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14 to 27B models on aggregate English and German benchmarks while achieving the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison, including ones far larger in active parameters. Among fully open models, Soofi S obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC scale AI infrastructure operated by Deutsche Telekom in Munich. Soofi S will be released under highly permissive, open-access terms: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, data-construction artifacts are released under permissive licenses; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.

Mixture-of-Experts (MoE) language models scale model ability with sparsely activated experts, making this architecture a standard recipe for modern large models. However, sparse activation does not remove the deployment burden of storing and serving all experts, and the available deployment budget can vary substantially across devices, users, and workloads. Existing MoE compression methods are still largely fixed-budget, typically optimizing one compressed endpoint at each chosen target budget. We study a different setting: converting a large pretrained MoE LLM into a nested family of deployable subnetworks across budgets. Our method first ranks expert FFN channels by their importance, then lets each expert learn a discrete action to prune its channels. By gradually increasing cost pressure, a single action-training run exports a series of action masks from high to low budgets, each of which identifies a reliable smaller subnetwork nested in the ranked base model. Moreover, we use a single recovery fine-tune at a mid pruning budget (40%) to recover degraded model quality and transfer the recovered model to other unseen budgets. Overall, our framework surpasses recent MoE compression baselines. Specifically, on Qwen2-57B-A14B, our method retains ~99.8% of base performance while pruning 50% of routed expert parameters even without fine-tuning. For deployment, our pruned subnetworks deliver real memory reduction and throughput gains, and further support realtime online budget switching with kernel-level co-design.

With the increase in model parameters and training data, the instruction following and generalization capabilities of Large VisionLanguage Models (LVLMs) have been significantly improved. Based on the Mixture of Experts (MoE) architecture, LVLMs expand their parameter capacity while maintaining the inference cost. However, traditional MoE methods employ a Top-k static routing strategy, which fails to account for variations in the input and adaptively select the number of experts, resulting in suboptimal resource utilization. In this paper, we propose viewing token routing as an information encoding task, framing dynamic routing as a Minimum Description Length (MDL) problem in encoding By validating the connection between MDL and gating entropy in the MoE scenario, we introduce Gating Entropy-based Uncertainty-aware Adaptive Routing (GeMoE) for MoE. Unlike traditional static or heuristic-based dynamic routing methods, GeMoE explicitly models the trade-off between model complexity and performance. By using gating entropy to assess the complexity of tokens, GeMoE adaptively determines the number of experts each token should engage. On a wide range of backbones and benchmarks, our method achieves 99.5% average performance retention compared to the original static routing, while improving average expert activation sparsity by 36.5%.

Sparse Mixture-of-Experts (MoE) architectures enable scaling LLM parameters under a fixed inference budget by activating only a small subset of experts via top-$k$ routing. While this preserves causality and suits autoregressive language models, the discrete top-$k$ operator is not differentiable, forcing a fixed number of active experts per input and resulting in inefficient use of computation. We propose SoftMoE, which replaces discrete routing with a truncated soft top-$k$ LapSum relaxation, allowing gradient-based optimization of expert routing. We further parameterize the mean number of active experts per layer and impose a global budget constraint, enabling the model to learn how to allocate expert capacity across layers. SoftMoE remains fully compatible with autoregressive modeling and achieves performance comparable to or better than sparse MoE on language modeling and downstream tasks, while activating significantly fewer experts. Notably, the learned allocation is highly non-uniform, with later layers activating more experts. The source code is publicly available$^\dagger$.

Mixture-of-Experts (MoE) architectures efficiently scale Large Language Models (LLMs) by activating only a small fraction of their experts per token, yet the full parameter count - dominated by the expert parameters - must be held in training and inference memory. To address this, we introduce Expert Tying, an architectural modification that shares expert parameters across consecutive transformer layers while preserving independent, layer-wise routing and attention. We evaluate this approach across common, state-of-the-art architectures, including OLMoE, Qwen3, and DeepSeek-style MoEs. Our pretraining experiments demonstrate that tying experts can reduce memory footprint by almost 2x at virtually no degradation in perplexity or downstream quality. By exploiting the parameter redundancy inherent in MoE pathways, our method provides a highly favorable compute-to-memory trade-off, advancing efficient training and scaling of next-generation LLMs.

Mixture-of-Experts Multimodal Large Language Models (MoE-MLLMs) offer remarkable performance but incur prohibitive GPU memory costs, making compression essential. Among PTQ methods, expert-level mixed-precision quantization has proven effective for MoE-LLMs, yet suffers notable degradation on MoE-MLLMs due to two overlooked biases in expert importance estimation. (1) At the cross-modal level, the numerical dominance of vision tokens causes expert selection frequency to be dominated by vision tokens, masking experts that are critical to the text modality; (2) at the intra-vision level, the large proportion of redundant vision tokens further skew frequency statistics, obscuring experts critical for informative visual content. To bridge gaps, we propose MODE, a modality-decomposed expert-level mixed-precision quantization framework for MoE-MLLMs that decomposes expert selection frequency by modality, filters redundant vision tokens to obtain denoised visual frequency, and further evaluates quantization sensitivity per modality as a complementary signal to frequency-based estimation. These signals are integrated into an Integer Linear Programming formulation to assign per-expert bit-widths under a given budget. Extensive experiments show that MODE is particularly well-suited for MoE-MLLMs, limiting average performance loss to within 2.9% at W3A16, with larger gains at the extreme 2-bit setting.

Mixture-of-Experts (MoE) models activate only a sparse subset of experts per token, yet consecutive tokens frequently activate different experts -- causing constant weight swapping between slow storage and fast memory on edge devices. Existing remedies are either system-level (caching heuristics) or post-hoc (router fine-tuning), leaving the root cause unchanged during pretraining. We propose StickyMoE, a differentiable routing consistency loss that penalises abrupt expert switches between adjacent tokens, encouraging the router to maintain the same expert assignment across semantically coherent spans. StickyMoE requires no architectural changes, adds a single hyperparameter lambda, and unlike post-hoc methods, allows expert representations and routing decisions to co-adapt from the first training step. Experiments on small-scale MoE language models show that StickyMoE reduces the expert switch rate by up to 60% with less than 4% perplexity degradation, Pareto-dominating post-hoc fine-tuning on the quality-locality frontier. Routing temporal locality is most efficiently instilled at training time.

Knowledge injection aims to equip large language models (LLMs) with external, domain-specific, or time-sensitive knowledge. Existing approaches typically face a trade-off between flexibility and integration: retrieval-augmented generation keeps knowledge outside the model but only provides prompt-level augmentation, whereas post-training based methods encode new knowledge into shared parameters but may introduce catastrophic forgetting, knowledge conflict, and costly updates. In this paper, we propose Decoupled Mixture-of-Experts (DMoE), a modular architecture for parametric knowledge injection that decouples both experts and the router from the base model. DMoE converts external knowledge corpora into independently updatable expert modules and uses a lightweight uncertainty-aware router to activate relevant experts only when the base model lacks sufficient knowledge during generation. To support efficient auto-regressive inference, DMoE attaches experts only to the final-layer feed-forward network, preserving KV-cache reuse while enabling parameter-level knowledge augmentation. Experiments on knowledge-intensive benchmarks show that DMoE consistently improves answer quality over retrieval and adapter-based baselines.

While Speech Large Language Models (Speech-LLMs) have achieved strong performance on adult Automatic Speech Recognition (ASR), their effectiveness on child speech remains under-explored, and single models often struggle to handle diverse adult and child age groups simultaneously. This paper proposes a Mixture-of-Experts (MoE) Speech-LLM for unified ASR across adult and child speech spanning diverse environments and age groups. The framework employs a Classifier-based Domain Router (C-DR) with a coarse-to-fine strategy and integrates both a Mixture-of-Projectors (MoP) and a Mixture-of-LoRAs (MoL) to model domain-specific variations. To address routing uncertainty near domain boundaries, an Entropy-Aware Routing (EAR) mechanism is introduced to dynamically incorporate a shared expert. Experiments on public child corpora demonstrate consistent improvements over baselines while preserving adult ASR performance. To our knowledge, this is the first work leveraging Speech-LLMs for unified, multi-domain ASR encompassing both children and adults.

Machine unlearning is increasingly important for large language models, yet unlearning in Mixture-of-Experts (MoE) architectures remains underexplored. Unlike dense models, MoE architectures employ a router at each layer to assign each token to a sparse subset of experts. In this work, we observe that forget data often activates a small subset of experts disproportionately, while these experts may receive much weaker activation from retain data. This forget--retain routing mismatch can leave forget-critical experts under-regularized during unlearning. To address this, we propose \textbf{TRACE}, Targeted Routing-Aware Calibration of Experts, for MoE unlearning. TRACE first detects forget-critical experts from offline activation statistics, and then calibrates retain regularization by reweighting token-level retain losses so that each selected expert's retain-side activation frequency better matches its forget-side counterpart. Experiments on WMDP and MUSE-BOOKS across multiple MoE LLMs show that TRACE consistently improves the forget-utility trade-off, yielding a 9% relative utility improvement over the strongest baseline under comparable forgetting quality and the best performance on three out of four MUSE-BOOKS metrics.

Continual learning in Large Language Models (LLMs) is hindered by the plasticity-stability dilemma, where acquiring new capabilities often leads to catastrophic forgetting of previous knowledge. Existing methods typically treat parameters uniformly, failing to distinguish between specific task knowledge and shared capabilities. We introduce Mixture of Sparse Experts for Task Agnostic Continual Learning (SETA), a framework that resolves the plasticity-stability conflict through adaptive sparse subspace decomposition into task-specific expert modules. Unlike standard updates, where tasks compete for the same parameters, SETA separates knowledge into unique experts, designed to isolate task-specific patterns, and shared experts, responsible for capturing common features. This structure is maintained through adaptive elastic anchoring and a routing-aware regularization that jointly protect shared knowledge at both the weight and routing levels and enable a unified gating network to automatically retrieve the correct expert combination during inference. Extensive experiments across diverse domain-specific benchmarks demonstrate that SETA achieves competitive or superior overall performance relative to state-of-the-art continual learning baselines, with particularly strong retention of early-task knowledge and improved backward transfer on LLaMA-2 7B and Qwen3-4B.

This paper reports on training a hundred-billion-parameter sparse mixture of experts on a single eight-GPU node, end to end. LightningLM 0.1V is a recurrence-backbone language model family grown in four stages from a small dense seed, through a 5B and a 9B mixture of experts, to a 120B model with 460 routed experts under top-12 routing. Each larger model is grown from the trained weights of the smaller one; active parameters rise monotonically from 1.78B at the dense seed to 5.93B at 120B (about 5% of the 118.67B stored). The full lineage runs on single nodes, the larger stages at 8K context, reaching a released training loss of 1.78 at 120B scale. This is a systems and experience report. It is organized around three disciplines. Reversibility: a reversible recurrence stack reconstructs activations in the backward pass instead of storing them, holding activation memory flat as the model grows. State-preserving growth: each expansion (dense to MoE, shallow to deep, few experts to many) is given as a reproducible principle paired with the failure that results from getting it wrong; several failures are silent. Single-node economics: the 120B trains through TQP, a strategy of quantized base expert weights and trained low-rank adapters that carries optimizer state on 2.26B adapter parameters rather than 100B+ resident in routed experts, cutting expert-path optimizer state by a factor of ~45. What is new is the integration of known primitives, not any primitive in isolation: one grown lineage running end to end on a single node, documented at practitioner level, with per-domain held-out loss as evidence that targeted capabilities (multilingual Indic competence, code) were learned by construction. Model family, tokenizer, and training code are released.

Mixture-of-Experts large language models (LLMs) scale efficiently through sparse activation, yet their deployment is fundamentally constrained by the large static parameter footprint of experts. Existing compression approaches either remove entire experts, disrupting routing topology and harming performance, or rely on unstructured weight pruning with limited practical efficiency. To address the limitations, we propose TENP, a structured Trapezoidal ExpertNeuron Pruning framework. Using a few samples, we identify and retain important experts, while applying expert neuron pruning (ENP) to less important experts, reserving model parameters in a trapezoidal pattern from shallow to deep layers. When evaluating expert importance, we jointly consider both the magnitude of the expert output and its ability to change the direction of the input vector. For ENP, we measure each neuron's projected contribution to the expert output to identify and retain important neurons. We conduct extensive experiments on the Qwen and DeepSeek models. Under a routing expert sparsity of 40% and an average of 63.76% activated expert parameters, the DeepSeek model suffers only a 1-point drop in accuracy compared to the full-parameter model. Moreover, it outperforms the full-parameter model by 10% on code generation tasks.

Mixture-of-Experts (MoE) and looped architectures scale models along two orthogonal axes, namely parameter capacity and effective depth. However, mainstream looped architectures rely on dense backbones that couple parameter count with per-token FLOPs, which makes it impossible to isolate the effect of iterative computation under matched budgets. To this end, we present LoopMoE, a looped MoE language model that integrates sparse routing with iterative weight-shared computation through two designs. The first is IterAdaLN, which resolves weight-sharing symmetry via a modulation signal jointly conditioned on the iteration index and the per-token hidden state. The second is a capacity-balancing strategy that recovers the attention-to-FFN active parameter ratio of well-tuned non-looped references. Together, these designs enable the first strictly controlled, head-to-head evaluation of a looped MoE against a Vanilla MoE under identical total parameters, per-token FLOPs, and active sublayer ratios. At the 3B scale, LoopMoE outperforms the Vanilla MoE on 8 of 9 downstream benchmarks with an average improvement exceeding 1 point. At the 9B scale, LoopMoE continues to outperform the matched Vanilla MoE, indicating that the architectural gain persists at larger scale. Our work establishes a controlled synthesis of sparsity and recurrence, and suggests a promising direction for looped language models.

Leveraging multiple specialized LLMs can combine complementary strengths, but existing approaches trade adaptability for stability: routing commits prematurely, heuristic ensembling depends on fragile proxies, and parameter merging introduces interference. We propose DLLG (Dynamic Logit-Level Gating), a dynamic logit-level ensembling framework that learns token-level expert fusion from sparse response-level supervision. A lightweight gating module predicts step-wise fusion weights, linking trajectory-level correctness to generation without token-level labels or expert retraining. Across diverse reasoning and code benchmarks, DLLG consistently outperforms strong routing, heuristic ensembling, and parameter-merging baselines across model scales, highlighting learned logit-level fusion as a robust and scalable paradigm for integrating specialized experts.

Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse to respond to harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to effectively suppress refusal behavior, inducing response to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs and find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture. We then propose two expert-aware refusal steering methods that leverage refusal-specific expert routing patterns and expert-specific steering directions to suppress normal refusal behavior. We find that refusal behavior can be effectively steered based on the output of a single expert. Our results show that refusal signals captured by steering methods differ from expert routing behavior, suggesting a substantial role for attention in MoE refusal behavior.

Activation patching can identify a mixture-of-experts (MoE) block whose clean output restores a corrupted factual prediction. However, because the block output combines contributions from multiple routed experts, block-level rescue does not establish whether the recovery localizes to an individual expert or depends on the routed expert set. We study this question on single-token COUNTERFACT contrasts by corrupting subject-token embeddings, restoring clean block outputs, and then restoring clean-minus-noised expert updates under fixed routing. In Qwen3-30B-A3B-Base, a discovery sweep selects layer 44, and held-out analysis identifies L44E069 as a recurrent routed contributor with positive specificity over same-layer active experts. Its effect is fact-matched and improves true-token probability and rank, which explains part of the layer rescue. In Mixtral-8x7B-v0.1, the selected recurrent singleton is not specific; matched-size controls instead show specificity of the clean-routed top-2 set. These findings show that successful MoE-block restoration does not necessarily imply localization to a single expert.

Mixture-of-Experts (MoE) architectures scale Large Language Models (LLMs) efficiently, enabling greater capacity with reduced computational cost by dynamically routing inputs to relevant experts, yet introduce a critical vulnerability: Safety Sparsity, where safety capabilities concentrate in few experts, making them susceptible to adversarial bypassing. Meanwhile, conventional alignment methods uniformly adapt all parameters, ignoring their functional differences and inadvertently degrading performances. To address these challenges, we propose MESA (MoE Safety Alignment), a targeted alignment framework for MoE-based LLMs that strategically decentralizes safety responsibility to maximize coverage while minimizing interference with utility. Based on Optimal Transport (OT) theory, MESA operates through two mechanisms: (1) Expert Capacity Reallocation uses a transport cost matrix to distribute safety duties to the most cost-effective experts, and (2) Dynamic Routing Refinement constrains the router to precisely activate these decentralized modules. Experiments show that MESA achieves robust defensive performance against varied harmful benchmarks while preserving helpfulness. Code is available at https://github.com/lorraine021/MESA.

Mixture-of-Experts (MoE) has become the dominant architecture for frontier language models. To meet this demand, production frameworks have built optimized MoE training stacks over years of engineering effort. Yet evolving these stacks for new architectures and system optimizations remains expensive. With the rise of AI coding agents, they could automate parts of training-framework development and accelerate this evolution. But applying them to these existing frameworks carries hidden costs, invisible to today's throughput-only evaluations. We name this missing dimension agent-task efficiency (ATE): the cost of using coding agents to understand, operate, and extend a framework. Grounded in four agent-native design principles, we build PithTrain, a compact, agent-native MoE training framework. We further introduce ATE-Bench, covering real-world training-framework tasks. Our evaluation shows PithTrain matches the throughput of production frameworks, and on ATE-Bench, PithTrain enables higher agent-task efficiency, with up to 62% fewer Agent Turns and 64% less Active GPU Time.

We present Mellum 2, an open-weight 12B-parameter Mixture-of-Experts (MoE) language model with 2.5B active parameters per token. Mellum 2 is a general-purpose language model specialized in software engineering, spanning code generation and editing, debugging, multi-step reasoning, tool use and function calling, agentic coding, and conversational programming assistance, and it is the successor to the completion-focused 4B dense Mellum model. The architecture builds on the Mixture-of-Experts (64 experts, 8 active) and combines Grouped-Query Attention with 4 KV heads, Sliding Window Attention on three of every four layers, and a single Multi-Token Prediction head that doubles as both an auxiliary pre-training objective and a built-in draft model for speculative decoding; each choice was validated by ablation with inference efficiency on commodity GPUs as a design constraint. Pre-training spans approximately 10.6 trillion tokens through a three-phase curriculum that progressively shifts the mixture from diverse web data toward curated code and mathematical content, optimized with Muon under FP8 hybrid precision and a Warmup-Hold-Decay schedule with linear decay to zero. The pre-trained base is extended to a 128K context window via a layer-selective YaRN and then post-trained in two stages (supervised fine-tuning followed by RLVR), yielding two released variants: an Instruct model that answers directly and a Thinking model that emits an explicit reasoning trace before its final answer. Across code generation, math and reasoning, tool use, knowledge, and safety benchmarks, Mellum 2 is competitive with open-weight baselines in the 4B-14B range while running at the per-token compute of a 2.5B dense model. We release the base, instruct, and thinking checkpoints, together with this report on the architecture decisions, data pipeline, and training recipe behind them, under the Apache 2.0 license.

Mixture-of-Experts (MoE) LLMs rely on sparse, router-driven expert activation, yet how safety alignment interacts with routed expert specialization remains underexplored. A common intuition is that safety behavior may be controlled by routing harmful requests to distinct refusal-oriented experts. In this work, we provide empirical evidence for a different picture: routing patterns in aligned MoE LLMs are largely topic-driven, while safety behavior can be altered with little change to the model's intrinsic routing path. Motivated by this observation, we present RASET (Router-Agnostic Safety-critical Expert Tuning), a red-teaming framework that probes safety enforcement that is localized in a small subset of experts while preserving the model's intrinsic routing behavior. RASET identifies safety-critical experts via a contrastive routing-sensitivity criterion and applies parameter-efficient tuning only to the selected experts, minimizing semantic disruption relative to router-steering interventions. These results reveal a distinct MoE safety risk, highlighting the need for expert-aware alignment mechanisms.

Mixture-of-Experts (MoE) language models reduce per-token computation but still require storing and serving all experts, making deployment memory-intensive. Existing post-training compression methods mainly shrink this cost by pruning experts or merging their weights. We formulate post-training MoE compression as expert-pool consolidation: retaining a smaller set of pretrained experts as reusable prototypes and deterministically remapping each original expert reference to one selected prototype. This view separates the reduced expert pool from the reuse structure that represents the original expert slots, and allows prototype sharing within local layer scopes while preserving the original router interface. We propose ConMoE, a train-free prototype remapping framework that selects retained experts using calibration-based contribution and replaceability signals, then redirects original expert calls to the selected prototypes without weight updates or post-compression fine-tuning. Experiments on three pretrained MoE language models show that ConMoE matches or outperforms strong pruning and merging baselines in several settings, achieving the best average score on deepseek-moe-16b-base at both 25% and 50% routed-expert reduction, while remaining competitive on Qwen3-30B-A3B and OLMoE-1B-7B-0125. Ablations indicate that deterministic reassignment is the most stable component, whereas broader cross-layer sharing and post-hoc weight fusion are model-dependent.

Mixture-of-Experts (MoE) models have emerged as a dominant paradigm for efficient LLM scaling, yet adapting them to non-English downstream tasks remains challenging. Existing fine-tuning approaches treat MoE models as monolithic learners, ignoring the heterogeneous routing structure that develops during pretraining. We validate across multiple MoE models and downstream tasks that middle layers form a language-universal alignment zone where routing divergence strongly predicts per-language task performance gaps. Building on this observation, we propose RA-MoE (Routing-Aligned MoE Fine-Tuning), a three-stage framework that categorizes parallel task examples into a four-way taxonomy (cc/ci/ic/ii) based on correctness in English and the target language, identifies task-relevant experts in the middle layers, and augments standard SFT with a routing alignment loss that encourages target-language routing on ci-type examples to follow the English task-expert activation pattern. Experiments across three MoE models, three tasks, and six target languages demonstrate that RA-MoE consistently outperforms standard SFT and strong baselines including Routing Steering and RISE, with the ci proportion of a task-language pair serving as a reliable predictor of alignment benefit.

Modern large language model (LLM) inference has progressively disaggregated to keep pace with growing model sizes and tight TTFT and TPOT service-level objectives: from chunked-prefill aggregation, to prefill-decode (P/D) disaggregation, and most recently to operator-level Attention-FFN Disaggregation (AFD). This trend is especially important for mixture-of-experts (MoE) models, where memory-bound attention, compute-intensive expert FFNs, and MoE dispatch/combine communication create distinct resource demands. AFD further exposes this heterogeneity by placing attention and MoE-FFN execution on separate GPU groups. Each level of disaggregation deepens the scheduling design space across workload characteristics, resource allocation, and interconnect topology, raising the central question: when does each level actually pay off? We systematically characterize this trade-off for MoE inference across realistic workloads spanning input/output sequence lengths, prefix-KV reuse, and per-user latency constraints. Using chunked-prefill and P/D disaggregation as baselines, we study the benefits and limits of AFD at scale through a framework that fuses on-device kernel measurements with high-fidelity network simulation. Under strict TTFT/TPOT SLOs, AFD sustains around 4k tokens/s of system throughput on DeepSeek-V3.2 across chat, coding, and agentic-coding workloads, where non-AFD deployments are infeasible. We distill concrete takeaways for jointly optimizing throughput and interactivity, including how to partition attention and FFN across GPUs as a function of workload and model architecture, providing design principles for current rack- and cluster-scale deployments as well as future disaggregated AI infrastructure.

Mixture-of-Experts (MoE) is now the dominant architecture for frontier language models, yet it requires all expert parameters to be loaded in memory, making it less preferable for memory-constrained deployment. Existing compression methods reduce the number of experts but the output remains an MoE model with the same fundamental limitation. We present the first systematic framework for converting a trained MoE into a standard fully dense architecture: experts are scored, selected, and grouped, then concatenated into a dense FFN and refined by knowledge distillation from the MoE teacher. We evaluate 7 scoring, 5 grouping, and 2 magnitude scaling methods across a range of selected expert counts on Qwen3-30B-A3B, yielding 350 configurations. We find that the choice of scoring method is the most impactful, with our novel diversity-aware scoring consistently outperforming prior methods on Qwen3-30B-A3B, DeepSeek-V2-Lite, and GPT-OSS-20B. Under a controlled comparison at matched parameter count, MoE-to-dense outperforms dense-to-dense pruning by +6.3 pp in average downstream accuracy after ~4B-token distillation at 1.6x faster training wall-clock speed.

Modern large language models (LLMs) achieve state-of-the-art machine translation performance, but they do so as broad generalists largely trained for many tasks and capabilities unrelated to translation. Thus, they are heavily overparameterized for this task, resulting in excessive memory and compute requirements. In this paper, we present a method for aggressively pruning experts from modern mixture-of-experts LLMs while incurring negligible degradation in translation quality. Our approach exploits expert specialization and the separability of multilingual capabilities in LLMs to identify experts irrelevant to translation. And because of the modular nature of MoEs, these can be easily pruned without any training. Without retraining, we are able to prune half of all experts with negligible degradation and 70% with only minor losses. With a very short SFT, we prune 75% of experts while recovering baseline performance, and in some settings remove nearly 90% while maintaining reasonable translation quality. Overall, our results show that translation requires only a fraction of the LLM, enabling substantial compression of the MoE blocks that contain over 90% of parameters.

Despite rapid progress in LLM-based code generation, existing models are predominantly trained on imperative languages, leaving functional programming languages (FPLs) such as Haskell, OCaml, and Scala chronically underexplored, with even frontier models performing substantially worse on FPLs. Fine-tuning is a natural remedy, but our experiments show that per-language fine-tuning fails to capture shared functional abstractions, while merged multi-language fine-tuning introduces cross-language interference. To address this, we introduce FPMoE, a lightweight, open-source code generation model built on a sparse Mixture-of-Experts (MoE) architecture with three language-specific routed experts (one each for Haskell, OCaml, and Scala) and a shared expert that captures cross-language functional patterns such as monadic reasoning and type-directed programming. This design resolves both failure modes simultaneously: dedicated experts eliminate interference, while the shared expert preserves abstractions that per-language models miss. On FPEval, FPMoE substantially outperforms fine-tuned baselines and, with only 3B active parameters, matches the performance of much larger models including DeepSeek-Coder-6.7B, Qwen2.5-Coder-14B-Instruct, and Qwen3-Coder-30B-A3B.

We present Laguna M.1 and Laguna XS.2, two Mixture-of-Experts foundation models built for long-horizon, agentic coding: M.1 has $225.8$B total parameters ($23.4$B activated per token) and XS.2 has $33.4$B total ($3$B activated). Both models were trained from scratch end-to-end inside the same internal system that we refer to as our Model Factory: a tightly-integrated stack of versioned data, training, evaluation, and inference components that turn model development into an industrial process. We describe the principles and design choices of the Model Factory and also detail the end-to-end training process of our models, throughout pre-training data and architecture, post-training stages, evaluation, and quantization. On agentic software engineering and terminal benchmarks (SWE-bench Verified, SWE-bench Multilingual, SWE-Bench Pro, and Terminal-Bench 2.0) M.1 and XS.2 are competitive with state-of-the-art open models in their respective weight classes. Laguna XS.2 weights are released under Apache~2.0 at https://huggingface.co/collections/poolside/laguna-xs2.

Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\times$ faster prefill and $2.2$-$3.4\times$ faster decode than the dense baseline MobileLLM-Pro.

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

While Large Language Models (LLMs) are commonly fine-tuned to handle domain-specific tasks before being applied to vertical applications, adapting them to complex scenarios with diverse specialized knowledge remains challenging. Meanwhile, Mixture-of-Experts (MoE) architecture has risen as a crucial paradigm for training LLMs, and some recent works have also incorporated MoE into Parameter-Efficient Fine-Tuning (PEFT) to propose the Mixture of Low-rank Experts (MoE-LoRA), to enhance the power of low-rank adapters for learning complicated knowledge. However, conventional gating mechanisms in MoE typically apply only a scalar reweighing to selected experts, thereby limiting their underlying capacity of representation and generalization. Motivated and enabled by the low-rank structures in MoE-LoRA, we propose RotMoLE, a specialized MoE framework for low-rank experts featuring an additional rotation gate. Beyond simple scaling, RotMoLE implements a rotation mechanism for each selected expert, enabling superior expert exploitation and specialization for learning diverse data, especially when expert candidates are limited. Empirical results on complex multi-task and multilingual training scenarios validate our effectiveness.

Mixture-of-Experts (MoE) architectures have become an increasingly important paradigm for scaling Large Language Models (LLMs). As MoE models are increasingly deployed in real-world services, safety auditing becomes necessary to verify whether these models produce or facilitate harmful behaviors during operation. However, existing content-based auditing methods typically require access to user prompts, model inputs, or generated outputs, potentially exposing sensitive user information and creating a fundamental tension between LLM safety and user privacy. On the other hand, we observe that, in MoE models, sparse expert routing maps different inputs to activate different expert-execution patterns, producing measurable footprints in low-level GPU execution telemetry. Inspired by this observation, we propose RouteScan, a non-intrusive auditing framework for detecting harmful behaviors through GPU-level expert routing telemetry. Specifically, RouteScan utilizes the number of active GPU threads allocated to expert modules during the prefilling phase as a discriminative micro-architectural fingerprint, and builds a lightweight detection pipeline that isolates cross-domain invariant risk indicators for the precise identification of malicious prompts. Comprehensive evaluations on open-source MoE LLMs with distinct routing designs demonstrate that RouteScan achieves strong generalization, with an AUROC exceeding 0.93 on unseen harmful domains and 0.96 under novel jailbreak wrappers. Moreover, empirical inversion tests show that the collected expert routing telemetry provides limited information for prompt reconstruction, suggesting a practical privacy advantage over content-based auditing methods.

Sparse mixture-of-experts (MoE) language models activate only a small subset of parameters for each token, making router behavior a central part of model computation. This paper studies routing behavior of Mixtral 8x7B-Instruct under benign and harmful prompts using two complementary signals: activation-based routing scores derived from expert selection frequencies and gradient-based scores derived from router-gate sensitivities. We analyze expert- and layer-level routing behavior and conduct expert-suppression interventions. The results show that activation-based expert usage is broad and long-tailed, whereas gradient-based importance is concentrated. At expert level, benign and harmful prompt groups remain close under both signals with modest separation. At layer level, activation-based routing is most selective around layers 8-15, while gradient-based importance is concentrated in final layers. Expert classification shows most experts are shared across benign and harmful prompts, though a limited subset shows clear group preference. Top-ranked expert sets show stronger benign-malicious overlap under gradient scores than activation scores, suggesting concentration on a common late-layer expert set. In intervention experiments, suppressing top five benign-dominant experts from activation scores reduces restricted responses from 24 to 14 over 100 prompts, while suppressing gradient-derived experts reduces them from 34 to 22 with fewer unintended reversals. Overall, safety-relevant routing in Mixtral is subtle, depth-dependent, and distributed rather than dominated by a fixed set of experts.

Mixture-of-Experts (MoE) large language models reduce per-token computation through sparse expert activation, but their deployment remains memory-intensive because all expert weights must be kept resident in memory. Existing MoE compression methods struggle in the ultra-low-bit regime: pruning irreversibly removes model capacity, while coarse-grained quantization fails to allocate bits according to heterogeneous expert and weight-direction importance. We propose BitsMoE, a spectral-energy-guided bit-allocation framework for MoE LLM quantization. BitsMoE decomposes each MoE layer by SVD into a shared basis and expert-specific spectral factors, retaining the shared basis without quantization to preserve common cross-expert structure and using the expert-specific factors as fine-grained quantization units. To determine the bit-width of each unit, BitsMoE formulates spectrum-wise mixed-precision quantization as an activation-aware reconstruction surrogate and solves an integer linear program that minimizes estimated reconstruction loss under a fixed bit budget. Experiments across multiple MoE LLMs show that BitsMoE substantially reduces downstream task accuracy degradation in ultra-low-bit regimes. Under 2-bit quantization on Qwen3-30B-A3B-Base, BitsMoE accelerates quantization by 12.3$\times$, improves average accuracy by 27.83 percentage points, and increases decoding speed by 1.76$\times$ over GPTQ. Our model and code are publicly available at https://github.com/zjiayu064/BitsMoE.

Mixture-of-Experts Large Language Models (MoE-LLMs) achieve strong performance but incur substantial memory overhead due to massive expert parameters. Mixed-precision quantization mitigates this cost by allocating expert-wise bit-widths based on their importance, approaching the accuracy-memory Pareto frontier and enabling extreme low-bit quantization. However, existing methods rely on layer-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing. In this work, we propose Global Expert-level Mixed-precision Quantization (GEMQ) to overcome these limitations via (1) a global linear-programming formulation that captures model-wide expert importance based on quantization error analysis, and (2) efficient router fine-tuning to adapt routing to quantized experts. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation. Source code is available at https://github.com/jndeng/GEMQ .

Expanding Large Language Models~(LLMs) to new languages is a costly endeavor, demanding extensive Continued Pre-Training~(CPT) and data-intensive alignment. While recent data-free merging techniques attempt to bypass alignment by fusing a multilingual CPT-enhanced model with its instruct counterpart, they are plagued by a critical trade-off: mitigating parameter conflicts to preserve original abilities inevitably dilutes new language acquisition, and vice-versa. To resolve this conflict, we introduce \method, which upcycles a dense model into a Mixture-of-Experts~(MoE) architecture, allocating different experts to different languages. Alignment ability is then transferred by grafting a MoE-expanded parameter delta~($Δ_{\text{post}}$) to the CPT-enhanced base model, bypassing the complex alignment phase. Experiments demonstrate \method's superiority even against baselines with similar FLOPs or number of parameters; it improves performance on expanded languages while effectively preserving original capabilities. We further show our approach is highly applicable across different models and Post-training deltas.

Catastrophic forgetting remains a major obstacle to continual learning in large language models (LLMs) and vision--language models (VLMs). Although Mixture-of-Experts (MoE) architectures offer an efficient path to scaling, existing LoRA-based MoE continual learning methods still face a fundamental trade-off: they either isolate experts too aggressively, limiting knowledge transfer across tasks, or allow task-specific updates to overwrite important existing parameters, leading to severe forgetting. To address this, we propose CP-MoE, a continual learning framework built around a transient expert that captures early task-specific updates and guides their integration into stable experts. CP-MoE introduces a consistency-preserving routing bias, which uses the transient expert to estimate representation similarity with stable experts and steer routing towards more compatible expert selection, and a transient expert-guided regularisation mechanism, which selectively protects important historical parameters during merging. Together, these components reduce parameter interference and forgetting while preserving cross-task knowledge transfer. We validate CP-MoE on both unimodal and multimodal continual learning benchmarks with LLM-based and VLM-based MoE models. On SuperNI benchmark, spanning diverse sequential language tasks, CP-MoE achieves state-of-the-art performance and stronger zero-shot transfer to unseen tasks. On VQA v2 dataset, it scales effectively to multimodal visual reasoning, consistently reduces forgetting, and outperforms strong MoE baselines.

Mixture-of-Experts (MoE) architectures enable efficient model scaling, yet expert routing behavior across underrepresented languages remains poorly understood. We analyze routing dynamics in two architecturally distinct MoE models -- a pure Transformer (Qwen3-30B-A3B) and a hybrid Mamba-Transformer (Nemotron-3-Nano-30B-A3B) -- using Hebrew as a morphologically rich, low-resource testbed. Both pre-trained models exhibit \emph{deep-layer routing collapse}: usage entropy drops sharply in final layers and tokens concentrate on a narrow expert subset, a pattern largely absent for English. Continual pre-training (CPT) on balanced bilingual data substantially corrects this imbalance, increasing entropy and shifting routing toward shared, language-agnostic experts; supervised fine-tuning (SFT) alone achieves less complete correction. Extending the analysis to Japanese reveals quantitatively consistent collapse signatures, providing cross-linguistic evidence that the phenomenon is a systematic consequence of pre-training underrepresentation rather than any language-intrinsic property. Routing improvements correlate with consistent downstream benchmark gains, positioning routing entropy and expert specialization as principled diagnostics for multilingual capacity in MoE systems.

Knowledge editing (KE) provides a lightweight alternative to repeated fine-tuning of LLMs. However, most existing KE methods target dense feed-forward layers, while modern LLMs increasingly adopt Mixture-of-Experts (MoE) architectures for their superior memory footprint and inference efficiency. This mismatch leaves a growing class of production models without principled editing tools. We propose a MEMIT-like framework for knowledge editing in MoE-based LLMs. Our method exploits the tensor structure of MoE layers to formulate the editing objective faithfully at the per expert level, and applies the Woodbury matrix identity to avoid materializing or inverting the full stacked matrix of expert weights. The resulting update reduces to inversions of fixed low-rank matrices and requires no additional backward passes. Empirically, our approach matches the editing quality of strong baselines on the main KE metrics while accelerating the editing procedure by up to 6x, owing to the batched MEMIT-style formulation and the low-dimensional inversions enabled by the Woodbury identity. These results show that closed-form, parameter-modifying KE can be extended efficiently beyond dense layers, opening a path toward scalable knowledge editing in modern sparse LLM architectures.

Large language models (LLMs) with mixture-of-experts (MoE) architectures achieve remarkable scalability by sparsely activating a subset of experts per token, yet their frequent expert switching creates memory bandwidth bottlenecks that compute-in-memory (CIM) architectures are well-suited to mitigate. However, analog CIM systems suffer from inherent hardware imperfections that perturb stored weights, and its negative impact on MoE-based LLMs in noisy CIM environments remains unexplored. In this work, we present the first systematic investigation of MoE-based LLMs under noise model calibrated with real chip measurements, revealing that hardware noise critically disrupts expert load balance and renders clean-trained routing decisions consistently suboptimal. Based on these findings, we propose ROMER, a post-training calibration framework that (1) replaces underactivated experts with high-frequency ones to restore load balance, and (2) recalibrates router logits via percentile-based normalization to stabilize routing under noise. Extensive experiments across multiple benchmarks demonstrate that ROMER achieves up to 58.6%, 58.8%, and 59.8% reduction in perplexity under real-chip noise conditions for DeepSeek-MoE, Qwen-MoE, and OLMoE, respectively, establishing its effectiveness and generalizability across diverse MoE architectures.