Meta releases Muse Glimmer, a 30B open-weights model that runs on a single consumer GPU

Open weights just landed on the desk. On August 10, 2026, Meta published Muse Glimmer, a 30-billion-parameter model that runs on a single consumer GPU.

What Meta shipped

Muse Glimmer is released under the Apache 2.0 license with the weights on Hugging Face. Meta built it for always-on local agent workflows: local agents, function calling, coding, and LLM-as-a-judge evaluation. It also carries a dedicated perception encoder, so it accepts interleaved text and images and can read screenshots, charts and documents, the inputs an agent actually encounters on a real desktop.

The number that matters is memory. Quantized, the model needs under 20 GB, which puts it inside the 24 to 32 GB envelope of a consumer graphics card. Meta says it tested on a MacBook M4-Max, an M5-Max and an RTX-5090. The runtime coverage is broad: Ollama, LM Studio and Unsloth locally, llama.cpp, ExecuTorch and MLX on the edge, vLLM and SGLang at scale.

Meta positions it against Gemma4-31B and Qwen3.6-27B on agentic, coding, multimodal, safety and reasoning benchmarks.

What comes next

Alexandr Wang said open weights for a version of Muse Spark 1.2, the bigger model in the family, are coming soon. The Wall Street Journal reports that as within the coming weeks. If that lands, Meta will have an open-weights ladder running from consumer hardware up to its larger tier.

Why local changes the calculus

The shift worth noticing is not the benchmark table, it is the trust boundary. A capable agentic model that never leaves the machine changes what a user is willing to point it at: private files, screenshots of internal tools, client work that cannot legally touch a third-party API. Every category of data that was off limits for hosted models becomes fair game for a local one.

That is also where the multimodal encoder earns its place. An agent that can read the screen, call functions and stay resident on a consumer GPU is the raw material for genuinely private automation. Whether Muse Glimmer is good enough to hold that role is what the head-to-head with Gemma4 and Qwen3.6 will decide.