Meta Releases Muse Glimmer: A 30-Billion-Parameter Open-Weight AI Model Built for Autonomous Agents
Summary
Meta launches Muse Glimmer, a powerful 30-billion-parameter open-weight AI model under Apache 2.0, designed for autonomous agents, capable of running on consumer hardware with 24GB VRAM, and outperforming key rivals on several agentic benchmarks.
Key Points
- Meta releases Muse Glimmer, a 30-billion-parameter open-weight AI model under the permissive Apache 2.0 license, marking the company's first fully open release since switching to the proprietary Muse Spark in April, with weights now available on Hugging Face.
- Designed specifically for autonomous AI agents, Muse Glimmer runs on consumer hardware with as little as 24GB of VRAM using 4-bit quantization, and employs DFlash speculative decoding to achieve up to 3.1x faster token generation speeds on Nvidia RTX 5090 GPUs.
- Glimmer leads competitors Gemma 4 and Qwen3.6-27B on several agentic benchmarks but does not dominate across all categories, entering a competitive local-model market where Chinese open-source models currently account for roughly 61% of tokens consumed on OpenRouter.