Meta Releases Muse Glimmer: A 30B Open-Source Model Excellelling in Agentic Tasks

kimmonismus · x · 2026-08-10

Meta has released Muse Glimmer, a highly capable 30B parameter model designed for local deployment. It achieves top results in 12 out of 24 benchmark rows, outperforming Gemma4-31B and Qwen3.6-27B overall.

Its standout feature is agentic performance, where it beats Qwen in tests like MCP Atlas, DeepSearch QA, and SWE-Bench Pro. However, Qwen still leads in OSWorld and most multimodal tasks, while Gemma holds the advantage in primary safety metrics.

The release is unusually open, providing ungated weights under Apache 2.0, including BF16 and quantized versions. Notably, the 4-bit model fits into roughly 17GB of memory with only a 1% average performance degradation across 15 benchmarks. This aligns with Zuckerberg's recent commitment to delivering accessible personal superintelligence.

Related event: Meta Releases Muse Glimmer 30B Open-Source Model for Local Agents(30 posts)→

Original post →

More from Models

Models channel →