Meta's Muse Glimmer: a 30B open agent model that runs on one GPU
Apache 2.0, multimodal, and agent-tuned — it beats Gemma4-31B and Qwen3.6-27B, ships GGUF quants under 20GB, and an open Muse Spark 1.2 is next.
Copy markdown
30B, Apache 2.0, and yours to download now
Muse Glimmer is a 30B open-weight agent model — a 2B vision encoder plus a 28B decoder — under a permissive Apache 2.0 license, so you can run, fine-tune, and ship it commercially. Weights are live on Hugging Face today.
It fits on the GPU you already have
Quantized to under 20GB, it runs on a single consumer card — Meta tested it on an RTX 5090 and MacBook M4/M5-Max. Local inference means no per-token API bill and no internet dependency for your agent.
Tuned to actually finish agent tasks
It scores 75.5 on MCP-Atlas and 51.2 on SWE-Bench Pro, with precise tool-call schemas, failure recovery, and a 'Controllable Effort' dial for reasoning depth — beating Gemma4-31B and Qwen3.6-27B on Meta's evals.
Multimodal, video included
A dedicated perception encoder takes interleaved text and images and reads video at 2 fps up to 96 frames, on a 32K context — enough to build a local agent that watches a screen recording or reads a diagram.
GGUF quants are already up
Pre-quantized GGUF builds from Unsloth and Meta's own repo drop straight into llama.cpp or a local server with no conversion step, and vLLM and Transformers paths ship alongside them.
The bigger model goes open next
Meta says an open-weight Muse Spark 1.2 — its most advanced model — is due in the coming weeks, as Zuckerberg pushes US policy to favor open-source AI. The gap between frontier and open keeps shrinking.