Meta Releases Muse Glimmer (3 min read)

Meta has released Muse Glimmer, optimized for local agent workflows, and open-sourced the model weights under the Apache 2.0 license. Designed with 30 billion parameters, this model can run on Macs or PCs equipped with a single consumer GPU.

Muse Glimmer aims to perform local agents, function calling, coding, and LLM-as-a-judge evaluations without cloud or internet connectivity. Model weights and developer documentation have been published on Hugging Face, with integrations for llama.cpp, MLX, and ExecuTorch also planned.

The model was developed using distillation techniques to transfer agentic reasoning capabilities from a large teacher model, along with inference optimizations.

• Pre-training: Logit distillation using outputs from Muse Spark
• Mid-training: Training on long-context and agent-centric data, including reasoning process data
• Post-training: Combination of supervised fine-tuning, on-policy distillation, and reinforcement learning

Muse Glimmer supports long-horizon task execution, precise tool calling, multi-step reasoning, tool failure recovery, text and image-based multimodal inputs, and agent orchestration compatibility. It also allows for adjustable reasoning intensity and was trained on data in over 100 languages.

It was evaluated on full-task benchmarks such as DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, with performance compared against similarly sized models like Gemma4-31B and Qwen3.6-27B.