We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that runs locally on consumer devices without requiring cloud infrastructure. The model accepts text and image inputs and produces text outputs, was trained on data from over a hundred languages, and is quantized to approximately 4-bit precision for efficient single-GPU inference. Muse Glimmer achieves state-of-the-art performance for its size class on agentic benchmarks, setting a new standard compared to the models of its size: Gemma4-31B and Qwen3.6-27B on tool use, multi-step planning, and agentic task completion. We release the model under the Apache 2.0 license to foster open-source development of capable, safe, local AI agents.
Figures & tables
Property
Value
Architecture
Dense Causal Transformer
Total parameters
∼29.6 B
Language model (transformer blocks)
∼25.16 B
Token embedding (untied)
∼1.34 B
Output projection (untied)
∼1.34 B
Perception encoder
∼1.8 B
Table 1: Architecture of Muse Glimmer.
Figure 1: Power law fits of bits-per-byte with respect to GPU-Hrs for the three ladders under consideration (GPU-Hrs is estimated based on final target model MFU). The dense ladder uses 550 TPP, while the MoE ladders use 2,800 TPP (1:8 sparsity) and 6,000 TPP (1:16 sparsity). The total parameter count for the target model in each of these options was chosen such that they can fit into 32 GB VRAM.
Capabilities Scorecard
Domain
Evaluation
Muse Glimmer (high)
Gemma4-31B (thinking)
Qwen3.6-27B (thinking)
Agentic & Coding
τ ³-Banking
23.5
15.1
16.7
WildClawBench
47.6
37.6
43.2
GAIA2
43.3
36.4
40.0
GDPval-AA
953
811
1141
SciCode
43.6
43.4
39.8
Table 2: Muse Glimmer capability scorecard. We compare Muse Glimmer at high reasoning strength against the competitive open models Gemma4-31B and Qwen3.6-27B across agentic, tool-use, reasoning, multimodal, memory, and privacy/safety evaluations. Best capability per row in bold . Higher is better unless marked ↓ . All benchmarks were screened for contamination.
Figure 2: Muse Glimmer capability radar chart. We compare Muse Glimmer at high reasoning strength against the competitive open models Gemma4-31B and Qwen3.6-27B
Figure 3: Memory and held-out validation quality of the evaluated PTQ methods. (left) Language-model weight size against mean ΔNLL across held-out text and code data. The chosen deployment profiles are highlighted. (center–right) Text and code ΔNLL for the same methods, ordered from lowest to highest within each panel. Lower values are better.
Model Variant
Average Accuracy (%)
Average Response Length
BF16
75.7
11070
EvoPress
75.6
11417
Greedy-KL
75.3
10754
Q4KM
74.8
10774
NVFP4/MXFP8
74.4
10988
Table 3: Downstream accuracy for quantization. Average accuracy and response length in tokens for different quantized variants across a mix of 15 downstream benchmarks.
Component
Setting
Draft layers
5
Block size
16
Attention
Sliding-window, 2048, all layers
Attention heads
32 query (matches target) / 8 KV (GQA)
Sequence length
128k
Hidden-feature layers
5, uniform over target: {1,13,25,37,49} of 52
Table 4: DFlash drafter configuration.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Capabilities Scorecard
Domain
Evaluation
Muse Glimmer (low)
Muse Glimmer (mid)
Muse Glimmer (high)
Muse Glimmer (xhigh)
Agentic & Coding
τ ³-Banking
–
–
23.5
26.0
WildClawBench
–
–
47.6
–
GAIA2
–
–
43.3
–
GDPval-AA
–
–
953
952
SciCode
–
–
43.6
39.2
Appendix
Table 5: Muse Glimmer capability scorecard. We compare Muse Glimmer at low , mid , high , and xhigh reasoning strength. Best capability per row in bold . Higher is better unless marked ↓ . All benchmarks were screened for contamination; “–” denotes not evaluated.
Type
Syntax
Semantics
Self
to=self
Internal reasoning and planning
Tool
to= namespace.function
Triggers tool execution
User
to=user
Final response; ends the turn
Appendix
Table 6: Recipient types in the Muse Glimmer routing protocol.
Language
Muse Glimmer (low)
Muse Glimmer (medium)
Muse Glimmer (high)
Muse Glimmer (xhigh)
Gemma4-31B (thinking)
Qwen3.6-27B (thinking)
eng_Latn (English)
73.7
73.7
73.7
73.9
43.2
56.6
spa_Latn (Spanish)
61.6
63.9
65.0
64.4
39.5
50.4
fra_Latn (French)
60.5
63.3
64.0
64.6
40.8
49.7
por_Latn (Portuguese)
62.7
64.6
64.9
65.3
40.8
48.3
ind_Latn (Indonesian)
59.2
62.0
62.1
62.1
39.0
48.2
ita_Latn (Italian)
58.8
60.7
61.8
62.3
39.7
46.6
Appendix
Table 7: Overall performance (pass@1, macro-averaged across the 4 capabilities) across the 10 languages of OmnilingualGAIA2 . We also report the English performance of GAIA2 as anchor. Target mean averages the 10 non-English languages, and Δ vs. English is its difference to the English row. Best result per row in bold.
Model
pass@1
Tokens/rollout
Muse Glimmer (low)
76.9
557
Muse Glimmer (medium)
79.3
1,180
Muse Glimmer (high)
80.1
2,453
Muse Glimmer (xhigh)
80.0
3,397
Gemma4-31B
82.5
2,671
Qwen3.6-27B
80.5
4,532
Appendix
Table 8: Accuracy vs. generation cost on MMLU-ProX. Tokens/rollout includes the reasoning trace.
Dataset
Direction
Language Pairs
Muse Glimmer
Gemma4-31B
Qwen3.6-27B
WMT24++
X → eng
55
55.7
57.2
55.7
WMT24++
eng → Y
55
51.0
52.2
48.8
FLORES+
X → eng
212
53.7
54.9
53.2
FLORES+
eng → Y
212
41.4
41.5
38.0
BOUQuET
X → eng
274
41.4
44.0
41.1
BOUQuET
eng → Y
274
31.7
33.3
29.7
Appendix
Table 9: Machine translation performance ( chrF++ ) of Muse Glimmer ( high ), Gemma4-31B and Qwen3.6-27B across three MT benchmarks and two aggregate directions. Overall averages the 972 FLORES+ and BOUQuET directions ( 212+212+274+274 ), the two benchmarks with the broadest language coverage.
Strong support ( n=88 )
afr_Latn (Afrikaans)
als_Latn (Tosk Albanian)
arb_Arab (Standard Arabic)
ast_Latn (Asturian)
azj_Latn (North Azerbaijani)
bel_Cyrl (Belarusian)
ben_Beng (Bangla)
bos_Latn (Bosnian)
bul_Cyrl (Bulgarian)
cat_Latn_stan1289 (Standard Catalan)
cat_Latn_vale1252 (Standard Valencian)
ceb_Latn (Cebuano)
ces_Latn (Czech)
cym_Latn (Welsh)
dan_Latn (Danish)
deu_Latn (German)
ekk_Latn (Standard Estonian)
ell_Grek (Greek)
Appendix
Table 10: Language support of Muse Glimmer, grouped by support tier.
We present Nanbeige4.2-3B, a compact general agentic model with 3B non-embedding parameters. It delivers strong performance across code-agent, office-agent, and complex tool-use tasks while maintaining highly competitive reasoning capabilities in mathematics, coding, and science. Nanbeige4.2-3B is pretrained from scratch on 28T tokens with a Looped Transformer that reuses the layer stack to increase capacity without adding parameters. For SFT data and trajectory construction, we expand the diversity of executable environments, task assets, and agentic scaffolds through real-world deployment and large-scale synthesis. Our RL pipeline applies mixed-mode RLHF over Think and Non-Think responses to improve overall model quality and reduce failure cases, length-controlled reasoning RL to balance accuracy and reasoning efficiency, and agentic RL with outcome and process rewards to stabilize long-horizon training. Extensive evaluations show that Nanbeige4.2-3B outperforms larger models, including Qwen3.5-9B and Gemma4-12B, across diverse agentic benchmarks while remaining competitive on reasoning and alignment tasks. Performance with OpenClaw further supports its use as a compact local personal assistant.
We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention language model. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.
Agentic language models dramatically expand the applications of AI yet little is publicly known about how to curate training data for broadly capable agents. Existing open efforts such as SWE-Smith, SERA, and Nemotron-Terminal typically target a single benchmark, leaving open the question of how to train models that generalize across diverse agentic tasks. The OpenThoughts-Agent (OT-Agent) project addresses this gap with a fully open data curation pipeline for training agentic models. We conduct more than 100 controlled ablation experiments to systematically investigate each stage of the pipeline, yielding insights on the importance of task sources and diversity. We then assemble a training set of 100K examples from our pipeline and fine-tune Qwen3-32B on this dataset, which yields an average accuracy of 44.8% across seven agentic benchmarks and a 3.9 percentage point improvement over the strongest existing open data agentic model (Nemotron-Terminal-32B, 40.9%). Moreover, our training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons. We publicly release our training sets, data pipeline, experimental data, and models at openthoughts.ai to support future open research on agentic model training.
Negin Raoof, Richard Zhuang, Marianna Nezhurina +47