Meta Unveils Muse Glimmer 30B: Open-Weights LLM Built for Local AI Agents
Published August 10, 2026
With Muse Glimmer 30B, Meta is launching an AI model designed not for classic chatbots, but for local AI agents — systems that work on tasks autonomously, call tools, analyze images and recover from errors, all on your own machine.
The model is released under the permissive Apache 2.0 license and is optimized to run on powerful PCs and Macs — fully local, even offline.
A 30-billion-parameter model made for agents
Muse Glimmer has around 29.6 billion parameters, a dedicated perception encoder for images (screenshots, diagrams, documents) and a context window of over 131,000 tokens. It was trained on data from more than 100 languages.
The real focus is agentic work: multi-step reasoning, function calling, failure recovery and long-horizon task execution. If a tool call fails, the model is meant to detect the error and retry instead of giving up. Meta explicitly targets agent frameworks such as OpenClaw and Hermes Agent.
Under 20 GB instead of 55+ GB
In full precision the model would need more than 55 GB of memory. Through 4-bit quantization, Meta shrinks it to under 20 GB — a roughly 17 GB variant targets GPUs with 24 GB of VRAM, with an average quality loss of only about one percent across 15 benchmarks. A 32 GB variant comes even closer to full precision.
DFlash: up to 3x faster decoding
With DFlash, Meta's speculative decoding technique, a small draft model proposes blocks of 16 tokens that the main model verifies in parallel. In Meta's own tests, speed on an RTX 5090 rose from 74.9 to 233.4 tokens per second (about 3.1x); on an Apple M5 Max from 26.6 to 50.2 tokens/s. AMD also published early results: up to 53 tokens/s on a Radeon AI PRO R9700.
Available via Hugging Face, Ollama and LM Studio
The weights — including quantized variants and the DFlash drafter — are on Hugging Face. Ollama already offers a 21 GB MLX version for Apple Silicon, and LM Studio supports the model as well. Via OpenAI- and Anthropic-compatible APIs, it plugs into existing local agent setups.
Strong on agent benchmarks — but not everywhere
Muse Glimmer scores 75.5 on the MCP Atlas benchmark (Gemma4-31B: 54.2, Qwen3.6-27B: 62.5), 74.6 on DeepSearch QA and 51.2 on SWE-Bench Pro. Qwen3.6-27B still leads on SWE-Bench Verified, TerminalBench and OSWorld-Verified — Muse Glimmer's strengths clearly lie in agent tasks, tool use and long workflows.
Why local AI agents matter
A permanently running agent needs access to far more personal data than an occasional chatbot. Running locally keeps that data on your own machine, avoids per-call API costs and works offline. Meta itself cautions against connecting the model to tools unprotected, recommending safeguards and human confirmation for irreversible actions.
Muse Glimmer 30B is a clear signal: models are increasingly optimized for concrete local agent workflows — and part of today's cloud-based AI agents could eventually move from the cloud to your own PC.
Sources
- Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device – Meta AI Research
- Muse-Glimmer-30B Model Card – Hugging Face / Meta
- Muse Glimmer – Ollama
- Run Meta Muse Glimmer 30B on AMD Ryzen AI Max and Radeon GPUs – AMD
