On-Device Language Model Inference

Deploying large language models on-device for always-on personal agents demands sustained inference from hardware tightly constrained in power, thermal envelope, and memory. We benchmark Qwen 2.5 1.5B (4-bit quantised) across four platforms: a Raspberry Pi 5 with Hailo-10H NPU, a Samsung Galaxy S24 Ultra, an iPhone 16 Pro, and a laptop NVIDIA RTX 4050 GPU. Using a fixed 258-token prompt over 20 warm-condition iterations per device, we measure throughput, latency, power, and thermal behaviour. For mobile platforms, thermal management supersedes peak compute as the primary constraint: the iPhone 16 Pro loses nearly half its throughput within two iterations, and the S24 Ultra suffers a hard OS-enforced GPU frequency floor that terminates inference entirely. On dedicated hardware, distinct constraints dominate: the RTX 4050 is bounded by its battery power ceiling, while the Hailo-10H is limited by on-module memory bandwidth. The RTX 4050 sustains 131.7 tok/s at 34.1 W; the Hailo-10H sustains 6.9 tok/s at under 2 W with near-zero variance, matching the RTX 4050 in energy proportionality at 19x lower throughput. Results should be interpreted as platform-level deployment characterisations for a single model and prompt type, reflecting hardware and software combined, rather than general claims about hardware capability alone.

Today, large language models have demonstrated their strengths in various tasks ranging from reasoning, code generation, and complex problem solving. However, this advancement comes with a high computational cost and memory requirements, making it challenging to deploy these models on edge devices to ensure real-time responses and data privacy. Quantization is one common approach to reducing memory use, but most methods apply it uniformly across all layers. This does not account for the fact that different layers may respond differently to reduced precision. Importantly, memory consumption and computational throughput are not necessarily aligned, further complicating deployment decisions. This paper proposes an adaptive mixed precision quantization mechanism that balances memory, latency, and accuracy in edge deployment under user-defined priorities. This is achieved by analyzing the layer-wise contribution and by inferring how different quantization types behave across the target hardware platform in order to assign the most suitable quantization type to each layer. This integration ensures that layer importance and the overall performance trade-offs are jointly respected in this design. Our work unlocks new configuration designs that uniform quantization cannot achieve, expanding the solution space to efficiently deploy the LLMs on resource-constrained devices.

Agricultural pest management increasingly relies on timely and accurate access to expert knowledge, yet high quality labeled data and continuous expert support remain limited, particularly for farmers operating in rural regions with unstable/no internet connectivity. At the same time, the rapid growth of AI and LLMs has created new opportunities to deliver practical decision support tools directly to end users in agriculture through compact and deployable systems. This work addresses (i) generating a structured insect information dataset, and (ii) adapting a lightweight LLM model ($\leq$ 7B) by fine tuning it for edge device uses in agricultural pest management. The textual data collection was done by reviewing and collecting information from available pest databases and published manuscripts on nine selected pest species. These structured reports were then reviewed and validated by a domain expert. From these reports, we constructed Q/A pairs to support model training and evaluation. A LoRA-based fine-tuning approach was applied to multiple lightweight LLMs and evaluated. Initial evaluation shows that Mistral 7B achieves an 88.9% pass rate on the domain-specific Q/A task, substantially outperforming Qwen 2.5 7B (63.9%), and LLaMA 3.1 8B (58.7%). Notably, Mistral demonstrates higher semantic alignment (embedding similarity: 0.865) despite lower lexical overlap (BLEU: 0.097), indicating that semantic understanding and robust reasoning are more predictive of task success than surface-level conformity in specialized domains. By combining expert organized data, well-structured Q/A pairs, semantic quality control, and efficient model adaptation, this work contributes towards providing support for farmer facing agricultural decision support tools and demonstrates the feasibility of deploying compact, high-performing language models for practical field-level pest management guidance.

Mobile input method editors (IMEs) are the primary interface for text input, yet they remain constrained to manual typing and struggle to produce personalized text. While lightweight large language models (LLMs) make on-device auxiliary generation feasible, enabling deeply personalized, privacy-preserving, and real-time generative IMEs poses fundamental challenges.To this end, we present HUOZIIME, a personalized on-device IME powered by LLM. We endow HUOZIIME with initial human-like prediction ability by post-training a base LLM on synthesized personalization data. Notably, a hierarchical memory mechanism is designed to continually capture and leverage user-specific input history. Furthermore, we perform systemic optimizations tailored to on-device LLMbased IME deployment, ensuring efficient and responsive operation under mobile constraints.Experiments demonstrate efficient on-device execution and high-fidelity memory-driven personalization. Code and package are available at https://github.com/Shan-HIT/HuoziIME.

Large language models (LLMs) with chain-of-thought reasoning achieve state-of-the-art performance across complex problem-solving tasks, but their verbose reasoning traces and large context requirements make them impractical for edge deployment. These challenges include high token generation costs, large KV-cache footprints, and inefficiencies when distilling reasoning capabilities into smaller models for mobile devices. Existing approaches often rely on distilling reasoning traces from larger models into smaller models, which are verbose and stylistically redundant, undesirable for on-device inference. In this work, we propose a lightweight approach to enable reasoning in small LLMs using LoRA adapters combined with supervised fine-tuning. We further introduce budget forcing via reinforcement learning on these adapters, significantly reducing response length with minimal accuracy loss. To address memory-bound decoding, we exploit parallel test-time scaling, improving accuracy at minor latency increase. Finally, we present a dynamic adapter-switching mechanism that activates reasoning only when needed and a KV-cache sharing strategy during prompt encoding, reducing time-to-first-token for on-device inference. Experiments on Qwen2.5-7B demonstrate that our method achieves efficient, accurate reasoning under strict resource constraints, making LLM reasoning practical for mobile scenarios. Videos demonstrating our solution running on mobile devices are available on our project page.

Realizing personalized intelligence faces a core dilemma: sending user history to centralized large language models raises privacy concerns, while on-device small language models lack the reasoning capacity required for high-quality generation. Our pilot study shows that purely local enhancements remain insufficient to reliably bridge this gap. We therefore propose SpecSteer, an asymmetric collaborative inference framework that synergizes private on-device context with cloud-scale reasoning. SpecSteer casts collaboration as Bayesian knowledge fusion and repurposes speculative decoding as a distributed alignment protocol, yielding a Draft--Verify--Recover pipeline: the on-device model drafts personalized sequences; the cloud validates via a ratio-based mechanism that decouples reasoning verification from private context, filtering logical flaws without accessing raw user context; upon rejection, a steering recovery injects local intent during correction. Experiments demonstrate that SpecSteer successfully closes the reasoning gap and achieves superior personalized generation performance, while delivering a 2.36x speedup over standard baselines.

Small language models (sLLMs) are increasingly deployed on-device, where imperfect user prompts--typos, unclear intent, or missing context--can trigger factual errors and hallucinations. Existing automatic prompt optimization (APO) methods were designed for large cloud LLMs and rely on search that often produces long, structured instructions; when executed under an on-device constraint where the same small model must act as optimizer and solver, these pipelines can waste context and even hurt accuracy. We propose POaaS, a minimal-edit prompt optimization layer that routes each query to lightweight specialists (Cleaner, Paraphraser, Fact-Adder) and merges their outputs under strict drift and length constraints, with a conservative skip policy for well-formed prompts. Under a strict fixed-model setting with Llama-3.2-3B-Instruct and Llama-3.1-8B-Instruct, POaaS improves both task accuracy and factuality while representative APO baselines degrade them, and POaaS recovers up to +7.4% under token deletion and mixup. Overall, per-query conservative optimization is a practical alternative to search-heavy APO for on-device sLLMs.

Real-time AI experiences call for on-device large language models (OD-LLMs) optimized for efficient deployment on resource-constrained hardware. The most useful OD-LLMs produce near-real-time responses and exhibit broad hardware compatibility, maximizing user reach. We present a methodology for designing such models using hardware-in-the-loop architecture search under mobile latency constraints. This system is amenable to industry-scale deployment: it generates models deployable without custom kernels and compatible with standard mobile runtimes like Executorch. Our methodology avoids specialized attention mechanisms and instead uses attention skipping for long-context acceleration. Our approach jointly optimizes model architecture (layers, dimensions) and attention pattern. To efficiently evaluate candidates, we treat each as a pruned version of a pretrained backbone with inherited weights, thereby achieving high accuracy with minimal continued pretraining. We leverage the low cost of latency evaluation in a staged process: learning an accurate latency model first, then searching for the Pareto-frontier across latency and quality. This yields MobileLLM-Flash, a family of foundation models (350M, 650M, 1.4B) for efficient on-device use with strong capabilities, supporting up to 8k context length. MobileLLM-Flash delivers up to 1.8x and 1.6x faster prefill and decode on mobile CPUs with comparable or superior quality. Our analysis of Pareto-frontier design choices offers actionable principles for OD-LLM design.

Device-side Large Language Models (LLMs) have witnessed explosive growth, offering higher privacy and availability compared to cloud-side LLMs. During LLM inference, both model weights and user data are valuable, and attackers may even compromise the OS kernel to steal them. ARM TrustZone is the de facto hardware-based isolation technology on mobile devices, used to protect sensitive applications from a compromised OS. However, protecting LLM inference with TrustZone incurs significant overhead due to its inflexible isolation of memory and the NPU. To address these challenges, this paper introduces FlexServe, a fast and secure LLM serving system for mobile devices. It first introduces a Flexible Resource Isolation mechanism to construct Flexible Secure Memory (Flex-Mem) and Flexible Secure NPU (Flex-NPU). Both memory pages and the NPU can be efficiently switched between unprotected and protected modes. Based on these mechanisms, FlexServe designs a fast and secure LLM inference framework within TrustZone's secure world. The LLM-Aware Memory Management and Secure Inference Pipeline are introduced to accelerate inference. A Multi-Model Scheduler is proposed to optimize multi-model workflows. We implement a prototype of FlexServe and compare it with two TrustZone-based strawman designs. The results show that FlexServe achieves an average $10.05\times$ speedup in Time to First Token (TTFT) compared to the strawman, and an average $2.44\times$ TTFT speedup compared to an optimized strawman with pipeline and secure NPU enabled. For multi-model agent workflows, the end-to-end speedup is up to $24.30\times$ and $4.05\times$ compared to the strawman and optimized strawman, respectively.

Immediate feedback is a foundational requirement of effective AI-mediated learning, yet the energy and latency costs of delivering it remain largely unexamined. This study investigates the latency-energy-learning trade-off in AI tutoring through an empirical comparison of two on-device inference configurations of Microsoft Phi-3 Mini (4k-instruct) on an NVIDIA T4 GPU: full-precision FP16 and 4-bit NormalFloat (NF4) quantisation. Both were evaluated under KV-cache-enabled inference across 500 educational prompts spanning five secondary school subject domains. Pedagogical quality was assessed for each of the 1000 generated responses by a hybrid panel of 10 Cambridge International teachers and three frontier AI systems using a four-dimension rubric. We introduce Learning-per-Watt (LpW), a novel metric quantifying pedagogical value per unit of energy over the learner's waiting window. Under realistic deployment, NF4 achieves lower per-inference energy than FP16 (329 J vs. 369 J) but higher latency (13.4 s vs. 9.2 s), yielding a modest FP16 advantage in LpW of 1.33x at a quality difference of 0.19 points. Under cache-disabled inference -- used in offline evaluation but absent from real deployments -- the gap widens to 7.4x, overstating the FP16 advantage by more than fivefold. Quantisation efficiency is hardware-dependent and inference-regime dependent, with significant implications for equitable AI tutoring deployment in low-resource settings.

Since local LLM inference on resource-constrained edge devices imposes a severe performance bottleneck, this paper proposes distributed prompt caching to enhance inference performance by cooperatively sharing intermediate processing states across multiple low-end edge devices. To fully utilize prompt similarity, our distributed caching mechanism also supports partial matching. As this approach introduces communication overhead associated with state sharing over a wireless network, we introduce a Bloom-filter-based data structure, referred to as a catalog, to determine whether a remote server possesses the desired internal states, thereby suppressing unnecessary communication. Experiments using the Gemma-3 270M model and the MMLU dataset on the Raspberry Pi Zero 2W platform demonstrate that the proposed approach reduces TTFT (Time to First Token) and TTLT (Time to Last Token) by 93.12% and 50.07% on average, respectively.

We investigate the feasibility of using Multimodal Large Language Models (MLLMs) for real-time online episodic memory question answering. While cloud offloading is common, it raises privacy and latency concerns for wearable assistants, hence we investigate implementation on the edge. We integrated streaming constraints into our question answering pipeline, which is structured into two asynchronous threads: a Descriptor Thread that continuously converts video into a lightweight textual memory, and a Question Answering (QA) Thread that reasons over the textual memory to answer queries. Experiments on the QAEgo4D-Closed benchmark analyze the performance of Multimodal Large Language Models (MLLMs) within strict resource boundaries, showing promising results also when compared to clound-based solutions. Specifically, an end-to-end configuration running on a consumer-grade 8GB GPU achieves 51.76% accuracy with a Time-To-First-Token (TTFT) of 0.41s. Scaling to a local enterprise-grade server yields 54.40% accuracy with a TTFT of 0.88s. In comparison, a cloud-based solution obtains an accuracy of 56.00%. These competitive results highlight the potential of edge-based solutions for privacy-preserving episodic memory retrieval.

The rapid proliferation of Large Language Models (LLMs) has revolutionized Natural Language Processing (NLP) but has simultaneously created a "resource divide." State-of-the-art legal intelligence systems typically rely on massive parameter counts (7B+) and cloud-based inference, rendering them inaccessible to practitioners in resource-constrained environments and posing significant data sovereignty risks. This paper introduces Quecto-V1, a domain-specific Small Language Model (SLM) engineered to democratize access to Indian legal intelligence. Built upon a custom configuration of the GPT-2 architecture (124 million parameters), Quecto-V1 was trained from scratch exclusively on a corpus of Indian statutes, including the Indian Penal Code (IPC), the Code of Criminal Procedure (CrPC), and the Constitution of India. Unlike generalist models, which prioritize broad world knowledge, our approach maximizes "lexical density" within the legal domain. Furthermore, we address the deployment bottleneck by applying post-training 8-bit quantization (GGUF format), compressing the model to a memory footprint of under 150 MB. Our empirical analysis demonstrates that Quecto-V1 achieves high fidelity in retrieving statutory definitions and penal provisions, outperforming general-purpose SLMs in domain-specific exact match tasks while running entirely offline on consumer-grade CPUs. We further present an ablation study showing that 8-bit quantization yields a 74% reduction in model size with less than 3.5% degradation in retrieval accuracy compared to full-precision baselines. These findings suggest that for specialized, high-stakes domains like law, domain-specific training coupled with aggressive quantization offers a viable, privacy-preserving alternative to monolithic cloud models.

Artificial intelligence and large language models (LLMs) are transforming educational technology by enabling conversational tutoring, personalised explanations, and inquiry-driven learning. However, most AI-based learning systems rely on continuous internet connectivity and cloud-based computation, limiting their use in bandwidth-constrained environments. This paper presents Arapai, an offline-first large language model architecture designed for AI-assisted learning in low-connectivity settings. The system performs all inference locally using quantized language models and incorporates hardware-aware model selection to enable deployment on low-specification, CPU-only devices. By removing dependence on cloud infrastructure, the system provides curriculum-aligned explanations and structured academic support through natural-language interaction. To support learners at different educational stages, the system includes adaptive response levels that generate explanations at varying levels of complexity: Simple English, Lower Secondary, Upper Secondary, and Technical. The system was evaluated with 120 students and 9 instructors from secondary and tertiary institutions under limited-connectivity conditions. Results indicate stable operation on legacy hardware, acceptable response times of 1-3 seconds for typical queries, and positive user perceptions of its effectiveness in supporting self-directed learning.

Large language model (LLM) queries are predominantly processed by frontier models in centralized cloud infrastructure. Demand growth strains this paradigm faster than providers can scale. Two advances create an opportunity to rethink it: small, local LMs (<=20B active parameters) now achieve competitive performance to frontier models on many tasks, and local accelerators (e.g., Apple M4 Max) can host these models at interactive latencies. This raises the question: can local inference viably redistribute demand from centralized infrastructure? This requires measuring both whether local LMs can accurately answer real-world queries and whether they can do so efficiently on power-constrained devices (e.g., laptops). We propose intelligence per watt (IPW), task accuracy per unit of power, as a unified metric for the capability and efficiency of local inference across model-accelerator configurations. We evaluate 20+ state-of-the-art local LMs, 8 hardware accelerators (local and cloud), and 1M real-world single-turn chat and reasoning queries. For each query, we measure accuracy (local LM win rate against frontier models), energy, latency, and power. We find three key results. First, local LMs successfully answer 88.7% of these queries, with accuracy varying by domain. Second, longitudinal analysis from 2023-2025 shows IPW improved 5.3x, driven by both algorithmic and accelerator advances, with locally-serviceable query coverage rising from 23.2% to 71.3%. Third, local accelerators achieve at least 1.4x lower IPW than cloud accelerators running identical models, revealing significant headroom for local accelerator optimization. These findings demonstrate that local inference can meaningfully redistribute demand from centralized infrastructure for a substantial subset of queries, with IPW serving as the critical metric for tracking this transition.

With the advancement of multimodal large language models (MLLMs), building GUI agent systems has become an increasingly promising direction--especially for mobile platforms, given their rich app ecosystems and intuitive touch interactions. Yet mobile GUI agents face a critical dilemma: truly on-device models (4B or smaller) lack sufficient performance, while capable models (starting from 7B) are either too large for mobile deployment or prohibitively costly (e.g., cloud-only closed-source MLLMs). To resolve this, we propose OpenPhone, a mobile GUI agent system that leverages device-cloud collaboration to tap the cost-efficiency of on device models and the high capability of cloud models, while avoiding their drawbacks. Specifically, OpenPhone enhances Qwen2.5-VL-3B via two-stage SFT->GRPO training on synthetic GUI data for strong decision-making, integrates an efficient long-reasoning and memory management mechanism to utilize historical interactions under tight resources, and defaults to on-device execution--only escalating challenging subtasks to the cloud via real-time complexity assessment. Experiments on the online AndroidLab benchmark and diverse apps show OpenPhone matches or nears larger models, with a significant reduction in cloud costs.

On-device deployment of Large Language Models (LLMs) frequently leverages Low-Rank Adapters (LoRAs) to support diverse downstream tasks under tight resource constraints. To address the limited storage capacity of mobile devices, recent works have explored model merging techniques to fuse multiple LoRAs into a single one. In practice, however, LoRAs are often delivered incrementally, as users request support for new tasks (e.g., novel problem types or languages). This scenario introduces a new challenge: on-device online continual merging, where the objective is to incorporate new LoRAs while preserving the performance on previously supported tasks. In this paper, we propose a data-free and computationally efficient strategy for selecting and merging LoRAs when a new one becomes available, assuming the device can store only a limited number of adapters. Extensive experiments across real-world tasks demonstrate the superiority of our approach compared to alternative strategies while adhering to the storage budget and compute limitations of on-device settings. The project page is available at: https://donaldssh.github.io/K-Merge.

On-device LLM inference is increasingly important for edge platforms but remains challenging due to high computational and memory demands and tight resource constraints. Modern applications often involve long prompts. Although compact models reduce linear-layer costs through weight compression, attention remains largely unoptimized and becomes the dominant bottleneck, especially during prefilling. Existing attention pruning methods primarily target highly parallel hardware (e.g., GPUs) and long context windows, limiting their suitability for edge scenarios. We present DeltaLLM, a training-free LLM framework based on algorithm-hardware-dataflow co-design to accelerate prefilling attention. DeltaLLM introduces a causal-attention-aware K-only delta formulation that enables Delta-Key reuse across query heads under Grouped-Query Attention (GQA). A context-aware hybrid strategy preserves exact attention near the diagonal while applying delta approximation to nonlocal regions. For score-value multiplication, we propose a vectorized structured sparse online softmax (VSSOS) that converts fine-grained spatial score sparsity into hardware-friendly column-level skipping. A specialized vector-processor architecture and unified sparse/dense attention dataflow exploit the resulting sparsity. Evaluated on BitNet-b1.58-2B-4T and LLaMA3.2-1B-Instruct, DeltaLLM achieves up to 60% Delta-Key sparsity and 63% structured attention-score sparsity with negligible average accuracy loss. On the KV260 FPGA board, it delivers a 1.62x speedup over TeLLMe-v2 while reducing DSP and BRAM usage. These results show that DeltaLLM supports efficient edge LLM deployment and integration with existing pipelines without fine-tuning.

On-device inference offers privacy, offline use, and instant response, but consumer hardware restricts large language models (LLMs) to low throughput and capability. To overcome this challenge, we present prima.cpp, a distributed on-device inference system that runs 30-70B LLMs on consumer home clusters with mixed CPUs/GPUs, insufficient RAM/VRAM, slow disks, Wi-Fi links, and heterogeneous OSs. We introduce pipelined-ring parallelism (PRP) to overlap disk I/O with compute and communication, and address the prefetch-release conflict in mmap-based offloading. We further propose Halda, a heterogeneity-aware scheduler that co-optimizes per-device CPU/GPU workloads and device selection under RAM/VRAM constraints. On four consumer home devices, a 70B model reaches 674 ms/token TPOT with <6% memory pressure, and a 32B model with speculative decoding achieves 26 tokens/s. Compared with llama.cpp, exo, and dllama, our proposed prima.cpp achieves 5-17 times lower TPOT, supports fine-grained model sizes from 8B to 70B, ensures broader cross-OS and quantization compatibility, and remains OOM-free, while also being Wi-Fi tolerant, privacy-preserving, and hardware-independent. The code is available at https://github.com/OpenCPIL/prima.cpp.

Deploying LLMs on edge devices presents serious technical challenges. Memory elasticity is crucial for edge devices with unified memory, where memory is shared and fluctuates dynamically. Existing solutions suffer from either poor transition granularity or high storage costs. We propose FlexQuant, a novel elasticity framework that generates an ensemble of quantized models, providing an elastic hosting solution with 31x more deployment options, 15x granularity improvement, and 10x storage reduction compared to SoTA methods. FlexQuant works with most quantization methods and creates a family of trade-off options under various storage limits through our pruning method. It brings great performance and flexibility to the edge deployment of LLMs.