Meta ships Muse Glimmer: 30B, Apache 2.0, built to run local agents on one GPU.
Muse Glimmer is a dense ~29.6B multimodal model from Meta Superintelligence Labs, open-weighted under Apache 2.0 and on Hugging Face today.1 It is trained around the agent loop — plan, call tools, check results, recover from failure — not chat. Quantized weights target under 20 GB so the model, KV cache, perception encoder and a DFlash speculative drafter fit a 24–32 GB envelope; Meta claims agentic reliability holds at 4-bit on that class of machine.12
Day-zero runtimes are already moving: transformers v5.15.0 adds the model; Ollama 0.32.7/0.32.8 adds MLX then NVIDIA/AMD paths; Hugging Face documents llama.cpp, vLLM and SGLang support landing with the weights.345 Integrations for Ollama, LM Studio and Unsloth are rolling this week rather than all live at publish time.1
weights and HF load path are real; treat consumer-GPU speed claims as vendor until you time a tool loop on your box. Pin a digest, run one agent scaffold you already trust, measure tokens/sec and tool-call success before you move anything off the cloud default.
Deeper — what is open, what is not
- Apache 2.0 covers full-precision BF16, 4-bit quants, the DFlash drafter and the ~1.8B perception encoder. Training data and training code are not released.2
- Context is stated at 131k+ tokens; knowledge cutoff 4 Jan 2026; text+image in, text out; 100+ languages; controllable effort via system prompt.2
- This is Meta's first fully open weights release since Muse Spark went proprietary. The licence is the news as much as the size.6
Sources
- research.meta.ai — Muse Glimmer · read
- huggingface — Muse-Glimmer-30B card · read
- github — transformers v5.15.0 · read
- github — ollama v0.32.8 · read
- huggingface blog — Muse Glimmer day-0 · read
- venturebeat — Apache 2.0 context · read