Muse Glimmer is Meta's 30B-parameter open model purpose-built for autonomous, agentic tasks on consumer hardware. It brings multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery together in a model that can run locally without cloud infrastructure.

Highlights

• Built for agents: Designed for long-horizon workflows, including planning, sequential tool calls, coding, and completing multi-turn tasks from start to finish.
• Reliable tool use and recovery: Follows function schemas across extended workflows and can diagnose failed or unexpected tool results before retrying.
• Multimodal input: A dedicated perception encoder processes interleaved text and images, including screenshots, charts, and documents.
• Controllable reasoning: Supports low, medium, high, and xhigh reasoning strengths to balance speed and quality.
• Long context: Supports a context length of 131,072 tokens or more.
• Multilingual: Trained on data spanning more than 100 languages.
• Agent scaffold compatibility: Designed to work with agentic orchestration patterns such as OpenClaw and Hermes Agent.

Optimized for local use

Muse Glimmer uses a dense causal transformer architecture with approximately 29.6 billion parameters, including a 1.8B-parameter vision encoder. Meta provides quantized variants targeting systems with 24 GB and 32 GB memory envelopes, making the model practical for high-end consumer PCs and workstations.

The release also includes a lightweight DFlash speculative-decoding drafter. It proposes blocks of tokens for the main model to verify in parallel, improving generation speed while preserving the main model's output.

Recommended uses

Muse Glimmer is particularly suited to:

• Local and autonomous AI agents
• Coding, debugging, and software-engineering workflows
• Tool use and structured function calling
• Screenshot, chart, document, and image understanding
• Synthetic data generation
• Evaluation and LLM-as-a-judge workflows

Reasoning strength

Reasoning strength can be selected in the system prompt:

Reasoning strength: high

Supported values are low, medium, high, and xhigh.

License

Muse Glimmer is released under the Apache 2.0 license.

For architecture details, benchmarks, intended-use guidance, and safety information, see the official Muse Glimmer model card.