Muse GlimmerOpen Weight

Muse Glimmer, released August 10, 2026, was the first open model from Meta Superintelligence Labs and the end of the closed stretch that began with Muse Spark in April: a 30B dense model distilled from Muse Spark 1.2, published on Hugging Face under Apache 2.0. Meta pitched it at always-on local agents rather than raw capability — small enough to run on a single consumer GPU or a Mac, multimodal through a dedicated perception encoder, trained on more than 100 languages, and tuned for the parts of agent work that break in practice: tool calls with exact schemas, plans held over long horizons, and recovery after a tool returns something unexpected. The engineering around it aimed at the same target, with 4-bit quantization bringing memory under 20GB and DFlash speculative decoding cutting generation time by 3.1x on an RTX 5090 and 1.8x on an M5 Max.

Meta benchmarked it against the open models in its size class, Gemma 4 31B and Qwen3.6 27B, and led on agentic tool use: 75.5% on MCP Atlas at release against 54.2% and 62.5% for the other two, plus 74.6% on DeepSearch QA and 47.6% on WildClawBench. It scored 76.0% on SWE-Bench Verified, 51.2% on SWE-Bench Pro, 83.5% on GPQA Diamond, and 94.7% on AIME 2026, while Qwen3.6 took the computer-use and long-form coding rows — 75.6% to Muse Glimmer's 65.9% on OSWorld-Verified, 60.7% to 51.7% on Terminal-Bench 2.1. Unlike the Muse Spark 1.2 launch, Meta published a methodology report naming every harness and judge, though most of the comparison numbers were still its own reproductions rather than each model's self-reported best. Meta said open weights for a version of Muse Spark 1.2 would follow.