Meta releases Muse Glimmer open-source AI model for local agents

Meta has released Muse Glimmer, a 30-billion-parameter open-weight AI model designed for autonomous agentic reasoning. Released under the Apache 2.0 license, the model is optimized for local execution on consumer-grade GPUs and workstations, eliminating the need for cloud subscriptions or API fees.

To enable high performance within limited memory budgets, Meta utilized a multi-stage training strategy involving logit distillation from its larger Muse Spark model, mid-training on complex reasoning traces, and post-training alignment. The model includes a 1.8B parameter perception encoder for processing multimodal inputs like screenshots and diagrams. Through dynamic quantization, the model's footprint can be reduced to approximately 17 GB to 20 GB, making it compatible with standard 24 GB to 32 GB hardware.

In agentic benchmarks, Muse Glimmer has outperformed competitors such as Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B in tasks requiring multi-step planning, tool use, and error recovery. Specifically, it scored 75.5 on the MCP Atlas tool-use benchmark and 51.2% on SWE-Bench Pro for software bug fixing. While it leads in autonomous reasoning, other models like Qwen3.6 still maintain advantages in specific areas such as OSWorld-Verified and TerminalBench 2.1.