Meta releases Muse Glimmer: 30B open-weight local AI agent model
Original: Meta Muse Glimmer – open weights 30B local coding model
Why This Matters
Frontier-grade agentic capability in a locally runnable open-weight model lowers the barrier for offline AI deployment.
Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weight model under Apache 2.0 license on August 10, 2026. Optimized for local agentic workflows, it runs on a Mac or PC with a single consumer GPU without requiring cloud connectivity.
Meta Superintelligence Labs introduced Muse Glimmer, a 30B-parameter model designed for always-on local agent workflows, releasing its weights on Hugging Face under the permissive Apache 2.0 license. The model targets use cases including local coding, function calling, and LLM-as-a-judge evaluation — all operable offline on consumer hardware such as a Mac or PC with a single GPU.
Training followed a three-phase process: pre-training via logit distillation from the larger teacher model Muse Spark; mid-training on longer-context, agent-heavy data with richer reasoning traces; and post-training combining supervised fine-tuning, on-policy distillation, and reinforcement learning across general, reasoning, coding, and agentic domains.
Muse Glimmer was evaluated on benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, measuring end-to-end task completion, reliable tool use, multi-step reasoning, and failure recovery. Optimized integrations with llama.cpp, MLX, and ExecuTorch are expected within days of release. The model was also assessed under Meta's Advanced AI Scaling Framework prior to open-weight release.