Muse Glimmer 30B: An On-Device Agentic Model

1 Introduction

• Muse Glimmer is a 30B model purpose-built for autonomous agentic tasks and distilled from Muse Spark to fit devices with 32GB of dedicated memory.Source: §1 · p. 1
• Muse Glimmer sets a new standard for its size class across tool use, multi-step reasoning, and task completion, performing strongly against Gemma4-31B and Qwen3.6-27B.Source: §1 · p. 1
• Targeted distillation is presented as producing a capable smaller model with fewer parameters and lower inference compute cost.Source: §1 · p. 1
• A dense architecture is presented as delivering high capabilities under the model-size memory constraint, compared with mixture-of-experts designs.Source: §1 · p. 1
• Combining 4-bit quantization with speculative decoding is reported to provide suitable edge-hardware speeds while preserving model quality.Source: §1 · p. 2
• The Apache 2.0 release includes full-precision and 4-bit weights plus a DFlash speculation head for community deployment, fine-tuning, and extension.Source: §1 · p. 2

2 Architecture

• The 30B causal language model uses a dedicated perception encoder to accept interleaved text and image tokens and produce text output.Table 1 summarizes the architecture.Source: §2 · p. 2, Table 1 · p. 2

2.1 Language model

• The language backbone has 52 transformer layers, 6656 hidden dimensions, and untied input and output embeddings, each contributing approximately 1.34B parameters.The backbone totals approximately 25.16B parameters excluding embeddings and output layers; its vocabulary has 202,048 tokens.Source: §2.1 · p. 2
• Attention uses 32 query heads and 2 key-value heads, with 2048-token local windows in a three-local/one-global layer pattern to bound KV-cache growth while retaining long-range information flow.RoPE is applied only to local layers; global layers rely on positional signals accumulated through surrounding local layers.Source: §2.1 · p. 2

2.2 Multimodal input and the perception encoder

• The perception encoder uses a NaViT-style ViT-G/14 and a pixel-shuffle, MLP adapter, and projection to map image tokens into the language model’s embedding space.A 448 × 448 image yields 256 output visual tokens; higher-resolution images can use up to 4096 tokens, and the encoder remains frozen during training.Source: §2.2 · p. 3
• Muse Glimmer can ingest video as frame sequences but is optimized for interleaved text-image understanding, including multi-image contexts, rather than video.Source: §2.2 · p. 3

2.3 Design for on-device efficiency

• Under a 32 GB VRAM ceiling, dense configurations were 2×–2.5× more compute-efficient than MoE counterparts, motivating the selected dense 30B design.The comparison held inference memory fixed and used progressively scaled architectural ladders.Source: §2.3 · p. 3
• Two simplified MLA variants underperformed standard GQA on long-context benchmarks such as RULER, so the design retained GQA.The variants matched either the baseline head count or its total KV footprint.Source: §2.3 · p. 4

• Autonomous agency involves chaining long-horizon reasoning, precise tool use, multimodal interpretation, and state retention across sessions.Source: §3 · p. 4
• Figure 2 compares Muse Glimmer at high reasoning strength with Gemma4-31B and Qwen3.6-27B in a capability radar chart.Source: Figure 2 · p. 5
• The evaluation covers multilingual agentic capabilities and non-agentic benchmarks for multilingual reasoning, reading comprehension, knowledge, and linguistic competence.Results are presented in Table 2 and Figure 2.Source: §3 · p. 5

4.1 Tokenizer

• The 202,048-token vocabulary contains 200,000 base tokens and 2,048 special tokens, with the base vocabulary trained on code, multilingual text, mathematics, and English.The mix is 35% code, 35% multilingual text, 15% mathematics, and 15% English.Source: §4.1 · p. 5
• A shared special-token budget unifies text, vision, reasoning, and agentic capabilities in one vocabulary space.Source: §4.1 · p. 5

4.2 Pre-training and Mid-training Stages

• Training proceeds through pre-training and two mid-training phases, using teacher-logit distillation for broad knowledge and reasoning before progressively adding domain-specific capabilities and longer context.Source: §4.2 · p. 5

• The base model is trained by logit distillation from Muse Spark, with top-256-logit co-distillation used during pre-training.Source: Pre-training · p. 5
• The co-distillation objective uses KL divergence for the first 95% of pre-training, then cross-entropy on ground-truth targets for the final 5% to address top-k truncation bias.Source: Pre-training · pp. 5–6
• Pre-training data spans web documents, code, mathematics, multilingual text, and images, with a knowledge cutoff of 2026-01-04.Source: Pre-training · p. 6
• A uniform average of the final checkpoint and three prior checkpoints spaced 600 steps apart improved perplexity evaluations over the single final checkpoint.The gains were larger on smoother benchmarks such as ML Papers and Multilingual Documents; the recipe followed a sweep of 84 checkpoints and merge strategies.Source: Pre-training · p. 6
• Mid-training Phase 1 adds synthetic reasoning traces and structured problem-solving demonstrations alongside organic data, using a lower learning rate to balance synthetic objectives against organic-perplexity regression.Source: Pre-training · p. 6
• Mid-training Phase 2 extends context from 32K to 128K tokens and adds tool-use and agentic trajectory data, using cross-entropy because teacher logits are less reliable under this distribution shift.The phase also uses purpose-generated agentic demonstrations and a learning rate below the peak pre-training rate.Source: Pre-training · p. 6

5 Post-training

• Post-training combines supervised fine-tuning, reinforcement learning, and on-policy distillation to convert the distilled base into an aligned agent.The stages target instruction following, reasoning, and agentic behavior.Source: §5 · p. 6

5.1 Supervised Fine-tuning (SFT)

• The SFT mixture covers reasoning, coding, mathematics, instruction following, tool use, and multi-step execution, prioritizing formatting, quality, and coverage over repeated exposure.Training uses approximately one epoch on each dataset to maintain entropy.Source: §5.1 · p. 6

5.2 Rollouts extraction

• Teacher trajectories from Muse Spark are filtered for reward and quality, then converted into conversational examples preserving model–environment interactions.Filtering removes evaluation metadata and malformed or unsupported tool interactions.Source: §5.2 · p. 7

5.3 On-Policy Distillation and Reinforcement Learning

• On-policy distillation trains on Muse Glimmer’s own rollouts with stepwise teacher feedback, addressing train–inference distribution shift and improving downstream RL outcomes over direct SFT-to-RL training.The student samples trajectories and the teacher evaluates its generated distributions step by step.Source: §5.3 · p. 7

5.4 Reasoning Modes

• Reasoning-strength training uses length-penalized subgroups and routes some prompts exclusively to higher strengths while others use the full spectrum.The full spectrum had a beneficial regularizing effect on chain-of-thought length, while higher-strength samples primarily drive capability acquisition.Source: §5.4 · p. 7

5.5 Trustworthy Local Agentic Workflows

• Privacy training teaches context-sensitive sharing by rewarding necessary information and penalizing oversharing across single- and multi-turn assistant tasks.The training draws on synthetic traces and includes tool-use settings.Source: §5.5 · p. 8
• Safety training targets indirect prompt injections, harmful requests, and conflicting instructions, while preserving completion of legitimate tasks.RL penalizes attacker-seeded actions and rewards successful task completion and recognition of untrusted content.Source: §5.5 · p. 8

6.1 Pre-training Data

• Pre-training spans web documents, code, STEM, multilingual text, and images, with domain filtering, quality scoring, deduplication, and benchmark decontamination.The reported knowledge cutoff is 2026-01-04.Source: §6.1 · p. 8

6.2 Synthetic Data

• Synthetic agent data uses simulated application environments and persona-based digital lives to generate constrained task scenarios with verifiable rubrics.Personas, tool states, and task sequences are constructed in stages, with tool states populated in dependency order.Source: §6.2 · pp. 8–9
• Post-processing removes failed or duplicated trajectories, masks undesirable actions, and filters hallucinations, false successes, and unvalidated outcomes.Source: §6.2 · p. 9

7 Inference Optimization

• Inference optimization focuses on low-batch edge devices, using weight-only quantization to reduce weight memory and decode data transfer alongside speculative decoding.Deployments target 24 or 32 GB of memory while reserving capacity for the unquantized KV cache, draft model, and perception encoder.Source: §7 · p. 9, §7.1 · p. 9

7.1 Quantization

• Quantization candidates target weight sizes below 20 GB and are selected using ∆NLL on held-out text and code; ∆NLL aligns reasonably with average accuracy across 15 downstream tasks.The authors report no significant chain-of-thought length variation relative to BF16 for their conservative quantization schemes.Source: §7.1 · p. 10, Table 3 · p. 11
• The winning EvoPress profile uses 418 independently exposed weight units and assigns 51, 130, and 237 units to Q4_K, Q5_K, and Q6_K, totaling 19.65 GB.It selects Q4_K for token embeddings and Q5_K for the output projection.Source: §7.1 · p. 10
• EvoPress has the lowest validation loss, while Q4KM has the smallest memory footprint; the paper selects them for 32 GB and 24 GB VRAM deployments, respectively.Figure 3 compares model weight size with mean held-out text-and-code ∆NLL and shows text and code ∆NLL separately; lower ∆NLL is better.Source: §7.1 · p. 10, Figure 3 · p. 11

7.2 Speculative Decoding

• DFlash uses a lightweight block-diffusion drafter that predicts a block of marginal probabilities in one forward pass; it is trained offline on target-model rollouts using D-PACE.Starting anchors for masked blocks are sampled at 1/16 of response tokens.Source: §7.2 · p. 11

8.1 Safety and Privacy Policies

• Risk coverage includes content safety, agentic risks such as irreversible actions and prompt injection, and catastrophic risks including chemical, biological, and cyber assessments.Mitigations combine training built into model weights with a lightweight inference-time safety mechanism.Source: §8.1 · p. 11, §8 · p. 11

8.2 Train-Time Mitigations

• Safety SFT adds examples for refusals, borderline prompts, tool-use boundaries, prompt-injection resistance, and permission handling; safety RL penalizes policy violations while rewarding helpful responses to legitimate requests.RL also iteratively evaluates checkpoints against content-safety evaluations.Source: §8.2 · pp. 11–12
• Dedicated synthetic training data embeds data-sensitivity recognition, minimization, and local-first execution into model weights as defense-in-depth beyond system prompts alone.Source: §8.2 · p. 12

On-device small language models

• Open-source Llama models helped establish the quantization, local-inference, and hardware-optimization infrastructure used for consumer deployment.The passage names GGUF, speculative decoding, and Apple Silicon optimizations.Source: §9 · p. 12
• The small-model landscape includes Phi-4, Gemma 4, and Qwen3, spanning dense and MoE designs, varied sizes, and multilingual or reasoning capabilities.Qwen3 supports 119 languages, while Gemma 4 uses quantization-aware training and KV-cache optimizations for memory and latency budgets.Source: §9 · p. 12

Teacher-student Distillation

• Prior distillation work transfers teacher reasoning to smaller students; Muse Glimmer’s recipe uses on-policy distillation from Muse Spark, which the authors found superior to RL-only training.Qwen3 uses student-generated rollouts with dense token-level KL supervision, while DeepSeek-R1 distilled reasoning traces into smaller students.Source: §9 · p. 12

Agentic LLM frameworks

• Agentic-model research covers interleaved reasoning and tool use, API-call training, and fine-tuning on multi-step trajectories; software agents also require sustained planning and failure recovery.The cited examples include ReAct, Toolformer, Gorilla, FireAct, Agent-FLAN, SWE-agent, and OpenHands.Source: §9 · p. 13

10 Conclusion

• Muse Glimmer surpasses comparably sized open models on agentic benchmarks and leads them by up to 27 and 23 pass@1 points on multilingual agentic evaluation.The report also states practical single-RTX 5090 speeds with 4-bit quantization and speculative decoding via a Dflash head, with minimal quality loss.Source: §10 · p. 13
• The report describes multimodal understanding of screenshots, documents, and charts, coverage of over a hundred language varieties, and release of all artifacts under Apache 2.0.Source: §10 · p. 13

A Benchmarks and Evaluation

• Table 5 compares capability scores at low, mid, high, and xhigh reasoning strengths; bold marks the best capability per row.Higher scores are better unless marked with a downward arrow; all benchmarks were screened for contamination.Source: Table 5 · p. 25

A.1 Agentic and Coding Task Completion

• The agentic and coding suite spans long-horizon tool use, desktop workflows, simulated service tasks, coding, and knowledge work.WildClawBench averages scores across 60 tasks and three runs; OSWorld-Verified measures GUI-only desktop control across 361 tasks, while GAIA2 tests tool use in stateful app environments. Other protocols cover TauBench, SWE-Bench, SciCode, and additional coding and work tasks.Source: §A.1 · pp. 26–27

A.2 Tool Use & Function Calling

• Tool-use evaluation includes sandboxed PinchBench work tasks and MCPAtlas queries across more than 20 MCP tool servers.PinchBench combines automated checks with an LLM judge and is treated as primarily diagnostic because it has 24 tasks, many with judge-graded free-form deliverables; MCPAtlas uses a 0.75 pass threshold and averages 500 tasks over four runs.Source: §A.2 · pp. 27–28

A.3 Reasoning & Knowledge

• Reasoning and knowledge are assessed with agentic web research, challenging mathematics and science questions, and visual abstract-reasoning puzzles.DeepSearchQA uses a judge-computed F1 score over 900 research questions; AIME 2026 uses 30 questions, GPQA Diamond uses 198, and ARC-AGI uses exact-match grid grading.Source: §A.3 · p. 28

A.4 Instruction Following & Steerability

• Instruction-following and steerability are evaluated through verifiable prompt constraints and agent tasks using mounted skill folders.IFBench contains 294 tasks and reports mean pass rate; SkillsBench uses 86 tasks with containerized environments and hidden pytest grading.Source: §A.4 · pp. 28–29

A.5 Multimodal Understanding

• Multimodal evaluation covers chart reasoning, complex documents, academic questions with images, and GUI element localization.The protocols include 1,000 CharXiv chart questions, 1,730 MMMU-Pro questions, 1,355 OmniDocBench prompts, and ScreenSpot screenshot-grounding tasks; ScreenSpot-Pro allows iterative cropping, unlike ScreenSpot v2.Source: §A.5 · pp. 29–30

A.6 Memory & Long-Context

• Long-context evaluation tests reasoning across documents and memory recall from conversations, without tools or scaffolding.AA-LCR uses 100 questions across document categories; BEAM-128K evaluates 400 conversations of 128K-token length across ten memory abilities.Source: §A.6 · p. 30

A.7 Multilingual

• Multilingual evaluation combines agentic tasks in target languages with text-only linguistic reasoning, comprehension, knowledge, and translation benchmarks.OmnilingualGAIA2 extends four GAIA2 capabilities to 10 languages, while Linguini tests reasoning across 75 mostly low-resource languages; other protocols include Belebele, MMLU-ProX, WMT24++, FLORES+, and BOUQuET.Source: §A.7 · pp. 30–31

A.8.1 Data Handling at Inference Time

• Inference-time privacy evaluation tests whether assistant responses expose sensitive information in simulated tool-use contexts.PrivacyLens uses 493 prompts and reports adjusted exposure conditioned on helpfulness; CIMemories covers synthetic profiles with over 100 attributes across 2,500 scenarios.Source: §A.8.1 · p. 31

A.8.2 Robustness to Prompt Injections

• Siren AgentDojo evaluates 97 benign tasks and 35 malicious tasks across four environments, totaling 949 prompt-injection scenarios, with Attack Success Rate and benign-task utility scores.An LLM attacker iteratively refines attacks for up to six iterations; deterministic rule-based grading measures attack success and benign task completion.Source: §A.8.2 · p. 32

B.3 Synthetic Data: Persona Generation Details

• Persona generation uses either sampled demographic and Big Five attributes or text-grounded passages to prompt coherent persona synthesis.The text-grounded mode produces plausible personas with explanations and probability scores, plus background on routine, relationships, and goals.Source: §B.3 · p. 35

B.4 Request Format

• Muse Glimmer supports tunable reasoning strength, with high as the default for agent workloads and xHigh reserved for offline or evaluation settings.Reasoning strength controls internal reasoning depth; xHigh provides the largest thinking budget for difficult long-horizon tasks where quality outweighs latency.Source: §B.4 · p. 35
• The request format routes messages through declared recipients and uses ATEM XML blocks for tool invocation.The recipient directive limits routing to explicitly listed recipients; self and user are always present, and routing to user ends the turn.Source: §B.4 · pp. 35–36

B.5 Agentic Data Format

• The agentic data format separates private reasoning, tool calls, and user-facing answers, and marks external tool observations as untrusted content.Tool calls appear as ATEM XML in message text and are mirrored in metadata.tool_calls.Source: §B.5 · p. 37

B.6 Multilingual Capabilities

• 137 language varieties are supported: 88 are classified as strong and 49 as usable, using machine translation as a proxy.The evaluation distinguishes linguistic competence from transfer of task competence outside English.Source: §B.6 · p. 38

B.6.1 Multilingual Agentic Capabilities

• Muse Glimmer leads Gemma4-31B by 21.0–23.2 points and Qwen3.6-27B by 13.4–15.7 points averaged across 10 languages, depending on reasoning strength.It is the top-performing model in all 10 languages at every tested strength.Source: §B.6.1 · p. 38, Table 7 · p. 39
• At high strength, Muse Glimmer falls from 73.7 English pass@1 to 50.0 in Japanese, a 23.7-point gap; the average drop across target languages is 13.3 points.The comparison averages three English rollouts; target-language average drops range from 13.0 to 15.3 points across reasoning strengths.Source: §B.6.1 · p. 38
• Across the three models, average capability performance follows the same order: search, execution, adaptability, then ambiguity.Qwen3.6-27B has an average English-to-target drop of 11.6 points and a worst drop of 20.0; Gemma4-31B's corresponding drops are 5.7 and 10.7 points.Source: §B.6.1 · p. 38
• Muse Glimmer gains 2.3 points from low to medium reasoning strength across target languages, then averages 60.4 at high and 60.5 at xhigh.Hindi and Japanese peak at medium; their search scores fall from 81.8 to 69.9 and 77.8 to 68.5, respectively, from medium to high.Source: §B.6.1 · p. 38

B.6.2 Non-Agentic Multilingual Capabilities

• Muse Glimmer scores 8.8 whole-problem exact match on Linguini, versus 3.8 for Gemma4-31B and 2.5 for Qwen3.6-27B, but trails Gemma4-31B on per-question exact match, 17.9 to 18.8.Linguini tests induction over unfamiliar language rules; its 75 languages lie almost entirely outside Muse Glimmer's supported set.Source: §B.6.2 · p. 39
• Muse Glimmer scores 87.8% on Belebele, between Qwen3.6-27B at 86.1% and Gemma4-31B at 88.6%; on MMLU-ProX it scores 80.1 mean pass@1.Excluding Wolof, Muse Glimmer's average English-to-language gap is 2.6 points, compared with 3.7 for Gemma4-31B and 5.2 for Qwen3.6-27B.Source: §B.6.2 · p. 39
• Qwen3.6-27B uses 4,532 tokens per rollout, 1.8× Muse Glimmer's high-mode token use, while excluding Wolof Gemma4-31B buys 1.5 points of accuracy for about 10% more tokens.The comparison is framed as a trade-off between accuracy and throughput.Source: §B.6.2 · pp. 39–40, Table 8 · p. 40
• Moving from high to xhigh changes each reported multilingual benchmark result by less than run-to-run noise while using 10–27% more generated tokens.The authors use high as their default operating point.Source: §B.6.2 · p. 40

B.6.3 Machine Translation Performance and Detailed Language Support

• Gemma4-31B leads the three translation benchmarks and aggregate directions, while Muse Glimmer is close behind and within 1.5 chrF++ points into English and 1.2 points out of English on WMT24++.Muse Glimmer's performance gap widens on BOUQuET, which has the broadest language coverage.Source: §B.6.3 · p. 40, Table 9 · p. 41
• Muse Glimmer supports 137 language varieties under the stated translation-based definition: 88 are classified as strong and 49 as usable.The tiers use chrF++ scores of at least 50 for strong support and 40–49 for usable support; the reported count is of language varieties, or languoids.Source: §B.6.3 · p. 40

B.7.1 Chart Understanding

• With tail pinning, student tail mass stays at 4.5 × 10^-4 and teacher mass at student top-K rises from 0.9934 to 0.9955; without it, tail mass grows to about 2.2 × 10^-3.The loss curves start around 0.009 and drop within the first one to two steps.Source: §B.7.1 · p. 42

B.7.2 Solve ODE

• For y'' − 4y' + 8y = 0, substituting y = e^rx gives roots 2 ± 2i and the solution e^2x(C1 cos 2x + C2 sin 2x).The solution is also given in amplitude-phase form.Source: §B.7.2 · p. 42