Meta Superintelligence Labs just open-sourced Muse Glimmer, a 30-billion-parameter model optimized for a very specific job: running AI agents locally, always on, without a cloud round-trip.
The weights are out today under Apache 2.0, which means this is a model you can actually own, fine-tune, and ship inside your own infrastructure.
Why local agents matter
Most agent deployments still assume you're online and paying per token. Running a model locally changes the economics: agents that work offline, keep their context on your machine, and cost nothing per request.
Muse Glimmer is designed for exactly that — it's small enough to run on a Mac or PC with a single consumer GPU, which makes it practical for:
- Local agents and function calling
- Coding assistance that never leaves your machine
- LLM-as-a-judge evaluation workflows
- Always-on automation that responds instantly
What it can do
The model was trained and evaluated across the capabilities that actually matter for autonomous agents:
- End-to-end task completion — long-horizon benchmarks like SWE-Bench and MCP-Atlas
- Reliable tool use — precise function calls with strict schemas across extended workflows
- Failure recovery — when a tool call fails, it diagnoses and retries instead of halting
- Multimodal input — interleaved text and images through a dedicated perception encoder
- Controllable effort — different reasoning strengths to balance quality and speed
- Multilingual — trained on more than 100 languages
It also holds its own against larger closed models in its size class: the benchmarks show strong agentic and safety scores versus Gemma4-31B and Qwen3.6-27B.
The practical side
Weights are live on Hugging Face (meta-models/Muse-Glimmer-30B), with optimized integrations for llama.cpp, MLX and ExecuTorch landing in the coming days. The promise is simple: download → working agent in minutes, no cloud account required.
For anyone building agentic systems, this is one of the more interesting open releases this year. Local-first agents just became practical.