Meta Superintelligence Labs released Muse Glimmer on August 10, 2026 — an open agentic model with 30 billion parameters, designed to run locally on a single GPU. The model weights ship under a permissive Apache 2.0 license, which allows commercial use. It marks Meta's return to an open-weights strategy after a run of closed releases.
Key takeaways
- 30 billion parameters, an agentic model optimized for local execution
- Apache 2.0 license, weights available on Hugging Face
- Runs on a Mac or PC with a single GPU, under 20 GB of memory when quantized to 4-bit
- DFlash speculative decoding: up to 3.1x faster generation on an RTX 5090
- Benchmarked against Gemma4-31B and Qwen3.6-27B on agentic, coding and multimodal tasks
A model for local agentic workflows
Muse Glimmer targets end-to-end agentic tasks: repeated tool use across long sequences, multi-step reasoning and recovery when a tool fails. The model takes text and images, supports more than 100 languages and lets you dial the reasoning effort. Meta also points to compatibility with orchestration patterns such as OpenClaw. The key difference from large closed models is the emphasis on running it yourself, without sending data to the cloud.
Consumer hardware and speed
The model fits under 20 GB of memory once quantized to 4-bit, letting it run on a single GPU in an ordinary computer. Meta added its own acceleration technique — DFlash, a variant of speculative decoding?speculative decoding: A generation speedup: a small model drafts several tokens at once and the main model only verifies them.. The company says it delivers up to 3.1x faster generation on an RTX 5090, plus 1.8x on the M5 Max and 1.5x on the M4 Max versus standard token-by-token generation. Integrations with llama.cpp, MLX and ExecuTorch are announced.
| Hardware | Generation speedup |
|---|---|
| RTX 5090 | up to 3.1× |
| M5 Max | up to 1.8× |
| M4 Max | up to 1.5× |
Openness against rivals
Unlike the closed agentic models from OpenAI or Anthropic, Meta bets on fully open weights under Apache 2.0. That license permits commercial use with no licensing fees. The model goes up against Gemma4-31B from Google and Qwen3.6-27B from Alibaba — rivals in the same size class. Meta claims strong results in its category but does not publish a numeric edge over each of them.
Why it matters
Opening the weights of an agentic model under Apache 2.0 lowers the barrier for companies that do not want to depend on the APIs of large providers. Local execution on a single GPU changes the cost and privacy math: data never leaves the machine, and per-token fees disappear. For Meta, it signals a return to the strategy that built an ecosystem around its earlier open models. The key question is quality — whether a model this size can hold reliability across long agentic tasks, where closed rivals still have an edge.
What's next
- Integrations with llama.cpp, MLX and ExecuTorch are announced but not shipped on launch day
- Documentation and deployment guidance are available at dev.meta.ai/docs/muse-glimmer
- No published numeric benchmark results against Gemma4-31B and Qwen3.6-27B for independent verification
Sources
- Meta AI Research — Introducing Muse Glimmer, an open agentic model
- VentureBeat — Meta returns to open source with Muse Glimmer, an Apache 2.0-licensed 30B model optimized for agents
- MarkTechPost — Meta AI Releases Muse Glimmer
