Muse-Glimmer-30B

Training Data

Data Sources

Multimodal content sourced from:

• Publicly available data
• Data provided by third parties
• Information from Meta's products and services
• Curated and enriched by external vendor networks and Meta personnel

Knowledge Cutoff

January 4, 2026

Language Coverage

Trained on data from more than 100 languages. The model has not been evaluated on all languages contained in the pre-training data; performance may degrade on languages outside the strongly supported set.

Training Approach

Muse Glimmer is distilled from Muse Spark, Meta's larger frontier model. The distillation process transfers capabilities for agentic tasks, multi-step reasoning, tool use, and multimodal understanding into the smaller 30B parameter model.

Safety Training Integration

Training data includes curated safety examples covering tool-use boundaries, prompt injection resistance, permission handling, data sensitivity recognition, and appropriate information flows. See the Train-Time Safety Mitigations snippet for details.

Released Artifacts

All artifacts are released under Apache 2.0:

Download Statistics

• Downloads (last month): 525,361
• HuggingFace followers: 1,750
• HuggingFace likes: 1.77k

Feedback

Questions, comments, and bug reports should be submitted via the HuggingFace page.

Considerations & Limitations

Muse Glimmer is a technology that carries known and unknown risks. Testing conducted to date has not, and could not, cover all scenarios.

Known Limitations

1. Inaccurate/biased responses — The model may produce inaccurate, biased, or objectionable responses to user prompts.
2. Multi-step reasoning errors — While optimized for agentic tasks, the model may still make errors in novel scenarios not well represented in training data.
3. No video optimization — The model is not explicitly optimized for video; video input is processed as individual frames.
4. Incomplete language coverage — The model has not been evaluated on all languages in pre-training data. Performance may degrade on languages outside the strongly supported set.
5. Quantization trade-offs — Quantized inference may show minor quality differences in edge cases compared to full-precision.
6. Age restriction — The model is not intended for use by individuals under 18. Deployers must assess and mitigate risks for systems that may be used by minors.

Responsible Use

• Developers should perform their own safety testing and tuning tailored to their specific applications.
• Usage Policy
• Implement additional guardrails (e.g. human-in-the-loop confirmation for irreversible actions) when deploying in agentic contexts where the model can take real-world actions.

Intended Use Cases

Muse Glimmer is intended for commercial and research use. The model is optimized for autonomous agentic tasks including:

Target Applications

• Local AI agents: Multi-step planning, sequential tool invocation, failure recovery, and long-horizon task execution running entirely on consumer devices.
• Coding agents: Writing, debugging, and resolving real-world software engineering tasks (e.g. SWE-Bench style workflows).
• Tool use and function calling: Reliable schema-based tool invocation across extended, multi-turn workflows.
• Multimodal reasoning: Interpreting screenshots, charts, documents, and images alongside conversation for agentic and information-rich environments.
• Synthetic data generation: Generating high-quality training data for downstream model development.
• LLM-as-a-judge evaluation: Serving as an evaluator for other models' outputs.

Out-of-Scope

• Use in any manner that violates applicable laws or regulations (including trade compliance laws).
• Use prohibited by the Apache 2.0 License terms.
• Audio input/output is not supported.
• Use by individuals under the age of 18.

Train-Time Safety Mitigations

Muse Glimmer incorporates three layers of safety training:

1. Safety SFT (Supervised Fine-Tuning)

Curated examples demonstrating correct safety behavior, including agentic safety scenarios covering:

• Tool-use boundaries
• Prompt injection resistance
• Permission handling

2. Safety RL (Reinforcement Learning)

Reinforcement learning with safety-specific reward signals that:

• Penalize policy violations
• Reward helpful responses to legitimate requests

3. Appropriate Information Flows

Principles of data sensitivity recognition, minimization, and local-first execution embedded directly into model weights through dedicated synthetic training data.

Evaluation Methodology

• Common use case evaluations measure safety risks for chatbot, visual QA, and similar applications
• Capability evaluations measure vulnerabilities inherent to specific model capabilities
• Adversarial evaluation datasets were built and used to evaluate systems composed of Muse Glimmer models plus safeguards
• Applications should be evaluated in context with dedicated evaluation datasets

Trust & Safety

Muse Glimmer should be deployed as part of an overall AI system with additional guardrails, not as a standalone endpoint.

Four Risk Axes

1. Content Safety — Standard alignment for refusal of harmful requests and calibrated responses to borderline prompts.
2. Agentic Risk — Policies for irreversible-action confirmation, data minimization, scaffold boundary respect, and indirect prompt-injection resistance.
3. Privacy (Appropriate Information Flows) — Respect for contextual integrity of information when interacting with third parties on an individual's behalf, inspired by CI theory.
4. Preparedness — Chemical & biological, cyber, and loss-of-control risks.

Preparedness Assessment

Muse Glimmer does not fall under the definition of “Frontier AI” in Meta's Advanced AI Scaling Framework (AAISF), since it is generally less capable than Muse Spark.

Cyber and Loss of Control risk levels are inferred from Muse Spark 1.0, which received the same designations.

Chem/Bio Benchmark Results

Security & Privacy Benchmarks

Source: Muse Spark Safety & Preparedness Report

Citation & References

Authors

Meta Superintelligence Lab

Release Date

August 2026

License

Apache 2.0

Key Papers

1. DFlash: Block Diffusion for Flash Speculative Decoding — arXiv:2602.06036 — Published Feb 5, 2026
2. Perception Encoder: The best visual embeddings are not at the output of the network — arXiv:2504.13181 — Published Apr 17, 2025

Reports

• Muse Glimmer Methodology Report
• Muse Spark Safety & Preparedness Report

Suggested Citation

@misc{muse-glimmer-30b,
  title={Muse Glimmer Model Card},
  author={Meta Superintelligence Lab},
  year={2026},
  url={https://huggingface.co/meta-models/Muse-Glimmer-30B}
}

Model Ecosystem

Model Tree

Muse Glimmer-30B serves as a base model for a growing ecosystem of community-derived variants.

Collection

Part of the Muse Glimmer Collection on HuggingFace: BF16 weights, GGUF k-quants, ExecuTorch builds, DFlash drafter (4 items).

Quantization Formats

Community quantizations are available for multiple inference engines:

• llama.cpp (GGUF K-quant)
• LM Studio
• Jan
• Ollama

Base Model

• Parent model: Muse Spark (distilled from)
• License: Apache 2.0
• HuggingFace: meta-models/Muse-Glimmer-30B
• Downloads (last month): 525,361
• Followers: 1,750

Vision-Language Capabilities

Muse Glimmer accepts interleaved text and image inputs through a dedicated perception encoder, enabling agents to interpret screenshots, charts, and documents alongside conversation.

Perception Encoder

• Architecture: ViT-G/14 (Vision Transformer, Giant patch size 14)
• Parameters: ~1.8B
• Layers: 50
• Width: 1536
• Patch size: 14
• State: Frozen at inference time
• Reference: Perception Encoder (arXiv:2504.13181)

Image Processing

• Max visual tokens per image: 4,096
• Input modalities: text + image
• Output modality: text

Multimodal Benchmarks

Use Cases

• Interpreting screenshots for agentic workflows
• Understanding charts and data visualizations
• Processing documents with mixed text/image content
• STEM diagram analysis

Limitations

• The model is not explicitly optimized for video; video input is processed as individual frames
• Audio input/output is not supported

Context Length

• Native context length: 131,072 tokens
• Extended context length: 131,072+ tokens (same as native)

Position Encoding

• Type: RoPE (Rotary Position Embedding)
• Theta (θ): 500,000
• Applied to: Local layers only (not global attention layers)

Attention Pattern

The model uses a repeating [Local, Local, Local, Global] attention pattern across its 52 layers. Local layers use a sliding window of 2,048 tokens with RoPE, while global layers attend to the full sequence. This hybrid pattern enables efficient long-context processing while maintaining full-sequence attention at regular intervals.

Long-Context Benchmarks

• Beam128K: 65.1 (best in class vs. Gemma4-31B: 58.2, Qwen3.6-27B: 63.0)

The Beam128K benchmark specifically evaluates the model's ability to maintain coherence and retrieve information across long contexts, where Muse Glimmer leads its size class.

Local Deployment & Quantization

Muse Glimmer was optimized for local deployment, designed to run at practical speeds on consumer hardware without sacrificing quality.

Quantization Options

Degradation measured as average accuracy change across 15 common benchmarks. The 4-bit quantization compresses the language model to under 20 GB, leaving headroom for KV cache, perception encoder, and speculative decoding drafter within a 24–32 GB envelope.

Speculative Decoding with DFlash

Muse Glimmer ships with a lightweight "drafter" model based on DFlash, a block-diffusion model that predicts entire blocks of 16 tokens in a single forward pass. The main model verifies these proposals in parallel.

DFlash Drafter Specifications:

Inference Speeds

Measurements: batch size 1, greedy decoding. M4/M5 via ExecuTorch, RTX 5090 via llama.cpp.

Supported Frameworks

• llama.cpp — Native support for GGUF K-quant formats on Nvidia GPUs
• ExecuTorch — Apple Silicon (M4/M5) inference
• Transformers — Full-precision research and fine-tuning

Thinking Control — Reasoning Effort Levels

Muse Glimmer supports flexible reasoning strength via the system prompt:

Reasoning strength: <value>

Supported Levels

The model dynamically adjusts its reasoning depth based on the specified level, trading off speed for quality. For agentic workflows involving tool use and multi-step planning, high is the recommended default.

Preserve Thinking

The model's thinking output can be preserved or discarded depending on the use case. For debugging and transparency, preserving the thinking trace is recommended. For production APIs where token efficiency matters, thinking output can be suppressed.

Best Practices for Muse Glimmer

Sampling Parameters

Reasoning Strength

Reasoning strength controls how much the model thinks before responding. It can be set via the system prompt as:

Reasoning strength: <value>

Supported levels: low / medium / high / xhigh

• Use high or xhigh for complex problem solving, coding, and agentic tasks.
• Use medium for general conversation.
• Use low for simple, fast responses where latency is critical.

Long-Context Processing

The model supports a context length of 131,072+ tokens. For optimal long-context performance, ensure the KV cache has sufficient memory headroom, particularly when using quantized weights.

Benchmark Results — Muse Glimmer-30B (High Reasoning)

General Agentic

Agentic Coding

Multimodal

General Capabilities and Reasoning

Security and Privacy

Source: Muse Glimmer Methodology Report

Architecture Details

Perception Encoder

~1.8B parameter ViT-G/14, 50 layers, width 1536, patch size 14. Frozen at inference time.

Reference: Perception Encoder (arXiv:2504.13181)

Muse Glimmer Key Capabilities

Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.

Core Capabilities

• End-to-end Agentic Task Completion — Strong success rates on DeepSearch QA, MCP-Atlas, 𝛕3-Bench and SWE-Bench
• Reliable Tool Use — Handles a wide range of function calls with precise schemas throughout extended workflows
• Multi-Step Reasoning — Chains reasoning over long horizons, sustaining coherent plans across complex workflows
• Failure Recovery — Diagnoses errors and retries when tool calls fail or return unexpected results
• Multimodal Input and Reasoning — Dedicated perception encoder accepts interleaved text and images
• Scaffold Compatibility — Works across OpenClaw, Hermes Agent, and other agentic orchestration patterns
• Controllable Effort — Supports different reasoning strengths to balance quality and speed
• Multilingual — Trained on data from more than 100 languages