Meta Releases Muse Glimmer for Local Agent Workflows

As Meta first reported, Meta Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter model optimized for local agentic workflows. Released under an Apache 2.0 license, the model's weights are available on Hugging Face.

According to Meta, the model is designed to operate on a Mac or PC with a single consumer GPU, targeting use cases such as local coding, function calling, and LLM-as-a-judge evaluation. To fit within a 24 GB or 32 GB memory envelope, Meta utilized quantization to compress weights to approximately 4-bit precision, reducing the model size to under 20 GB. The system also employs a DFlash-based "drafter" model for speculative decoding to increase generation speed.

Meta reports that Muse Glimmer was developed using a distillation recipe that transferred reasoning from a larger teacher model, Muse Spark. The training process involved three phases: pre-training via logit distillation, mid-training on agent-heavy data, and post-training using supervised fine-tuning and reinforcement learning.

In terms of performance, Meta states that Muse Glimmer shows strong success rates on benchmarks including SWE-Bench, MCP-Atlas, DeepSearch QA, and 𝛕-Bench. The company notes the model performs strongly for its size class when compared to Qwen3.6-27B and Gemma4-31B.

Developer integrations for ExecuTorch, MLX, and llama.cpp are expected in the coming days. The model supports multimodal input through a perception encoder and is compatible with OpenClaw orchestration patterns.