
Muse Glimmer
6.4/10Muse Glimmer is Meta's open-weight 30B-parameter agentic model, released under Apache 2.0 to run locally on a single consumer GPU for coding agents and tool use.
Data last updated November 2026 · Our methodology
Overview
Muse Glimmer is an open-weight language model from Meta Superintelligence Labs, released on August 10, 2026. It is a dense transformer of roughly 29.6 billion parameters, distilled from Meta's larger internal Muse Spark model. Meta built it for agent work rather than general chat: multi-step planning, tool and function calling, recovering when a step fails, and local coding agents. It takes interleaved text and images, such as screenshots, charts, and documents, and returns text. Its context window is about 131,000 tokens, its training data covers more than 100 languages, and you can set reasoning effort to low, medium, high, or xhigh. With about 4-bit quantization it fits on one 24 GB or 32 GB GPU, so it runs on a Mac or PC without a cloud connection. That suits teams that need data to stay on their own machines. The weights are free under Apache 2.0, with no commercial-use thresholds. Meta does not offer a hosted API, so costs depend on your own hardware or on third-party hosts such as NVIDIA's build.nvidia.com trial service, Together AI, or Fireworks AI. It runs on vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, and ExecuTorch. The main limitation is that this is a bare model, not a product. You get no admin console, no SSO, no vendor support, and no managed interface. Its benchmark claims against Gemma4-31B and Qwen3.6-27B come from Meta itself, and teams need engineering skill to deploy and operate it.
Our Rating
Who Muse Glimmer is best for
Best software for DevelopersGood fit
Meta built it for local coding agents and tool calling, and it runs on common developer runtimes.
The free Apache 2.0 weights remove licensing costs, but the team has to run its own infrastructure.
Less suited
Local deployment keeps data private, but there are no vendor enterprise controls or support.
Pros & Cons
Pros
- The Apache 2.0 license allows commercial use with no user or revenue thresholds.
- The quantized variants run on a single 24 GB or 32 GB consumer GPU without any cloud connection.
- It is built for agent work: tool calling, multi-step planning, and recovering from failures.
- It reads screenshots, charts, and documents and has a context window of about 131K tokens.
- It works with many runtimes, including vLLM, Ollama, LM Studio, llama.cpp, and MLX.
Cons
- Meta offers no hosted API, so you must host it yourself or use a third-party provider.
- It outputs text only and cannot generate images, video, or audio.
- The benchmark results come from Meta and have not been independently verified.
- There is no vendor support, admin console, or enterprise controls such as SSO.
- Full-precision weights need more than 55 GB of memory.
Key Features
Not included
- Image Generation
- Voice and Audio
- Video Generation
- Enterprise Controls
Muse Glimmer Pricing
Open weights (self-hosted)
- Context window of about 131K tokens
- DFlash speculative-decoding drafter
- Apache 2.0 license with no commercial thresholds
- BF16 weights and 4-bit variants for 24 GB and 32 GB GPUs
Frequently Asked Questions
The weights are free under Apache 2.0. Your costs are your own hardware or a third-party host's charges.
No. Meta does not host an API. NVIDIA offers a trial service on build.nvidia.com, and providers such as Together AI and Fireworks AI host the model.
The about 4-bit quantized variants target single GPUs with 24 GB or 32 GB of memory. Full precision needs more than 55 GB.
Yes. It accepts interleaved text and images, including screenshots, charts, and documents, but it only produces text.
Yes. The Apache 2.0 license sets no commercial-use thresholds.
Similar Tools
See all alternatives
Open-source Python framework from deepset for building production LLM applications, including RAG pipelines, agents and semantic search, with optional paid enterprise support.

Anthropic's business plans (Team and Enterprise) for Claude, giving organizations shared AI chat, coding, agents and connectors with central admin and billing.

Onyx, formerly Danswer, is an open-source AI assistant and enterprise search tool that lets staff search and chat across 50+ connected workplace apps with permissions respected.
Ranked near Muse Glimmer in AI Software
See Muse Glimmer in the listMuse Glimmer is #359 in AI Software by our score.





