Meta Releases Muse Glimmer 30B Open Agent Model

Meta Superintelligence Labs has published the open-weight technical report and Apache 2.0-licensed weights for Muse Glimmer 30B, a dense agentic model tailored for workstation deployment. Built around a 2B perception encoder and a 28B decoder, the model provides a context window of over 120K tokens. While full BF16 weights consume roughly 60 GB, 4-bit weight quantization compresses the footprint to under 20 GB with a reported 1.0% degradation, allowing local deployment on a single 24 GB GPU or unified-memory Mac.

Distilled from Meta's larger proprietary Muse Spark system through logit distillation, supervised fine-tuning, and agent-focused reinforcement learning, the model is engineered for multi-step tool execution. To address the computational overhead of iterative agent reasoning, Muse Glimmer pairs with a DFlash drafter model for speculative decoding. On an RTX 5090 GPU, this setup delivers a 3.1x speedup, accelerating throughput from 74.9 to 233.4 tokens per second.

In benchmark evaluations, Muse Glimmer led the MCP Atlas benchmark for Model Context Protocol tool use, outperforming Qwen3.6-27B by 13 points and Gemma4-31B by 21.3 points. It also recorded a 51.2 score on SWE-Bench Pro for repository-level software resolution, alongside evaluations on DeepSearch QA. Meta is distributing the model across multiple formats, including standard BF16 weights, GGUF k-quants, and ExecuTorch builds.

For enterprise practitioners and software developers, Muse Glimmer shifts the trade-offs of running autonomous agent pipelines locally. Development teams can execute long-horizon coding tasks, process multimodal inputs, and run model-as-judge evaluations without submitting private code bases to third-party cloud APIs. Supported by OpenClaw integration and capabilities across more than 100 languages, the permissive Apache 2.0 license provides a flexible foundation for self-hosted agent infrastructure.