GitHub radar
Meta Muse Glimmer: local agent model via Ollama
Meta's Muse Glimmer-30B — distilled from their internal flagship, scoring 94.7% on AIME 2026 and 76% on real coding tasks — is now available locally via Ollama.
Meta Muse Glimmer-30B is an open multimodal model distilled from Meta's internal flagship Muse Spark, tuned for local deployment on consumer hardware with 24–32 GB VRAM. It was built for always-on autonomous agents: tool use, code writing, failure recovery when tool calls go wrong, and multi-step reasoning across text and images. According to Meta's official benchmarks, it scores 94.7% on AIME 2026 (math olympiad), 76.0% on SWE-Bench Verified (real-world coding), and 75.5% on MCP Atlas (agent evaluation). The context window is 128,000 tokens. Licensed under Apache 2.0. Run locally via: ollama run muse-glimmer.
Why a vibe-coder should care
Muse Glimmer brings Meta's strongest reasoning to local hardware — a model that scores 94.7% on math olympiad problems and 76% on real coding tasks without any cloud dependency. Particularly useful for long agent workflows where you can't afford rate limits or per-token costs.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Meta Muse Glimmer locally: ollama run muse-glimmer — show me how to configure it as a persistent agent for long multi-step tasks
A Mac with 32 GB of RAM or more, or a GPU with 24 GB — the model takes 18 GB.
Open on Hugging Face▌ More finds