Summary & Key Takeaways ​

Meta introduces Muse Glimmer, a 30B open-weights model with an Apache 2.0 license. The model is specifically optimized for end-to-end agentic task completion. It demonstrates strong performance in reliable tool use and multi-step reasoning. Simon Willison tested Glimmer locally using LM Studio and his llm-coding-agent plugin. The 30B size is practical for local machines with 32GB+ RAM, leaving room for other applications. Glimmer also functions as a vision model, capable of describing images. Its clean Apache 2.0 license is a notable improvement over previous Meta model licenses. Benchmarks show strong success rates on full-task benchmarks like SWE-Bench.

Our Commentary ​

Okay, Meta is back in the open-weights game with a proper Apache 2.0 license. That's a huge win for the community. We've been waiting for something like this, especially for local agentic workflows. The 30B size feels like a sweet spot for many of our machines. I'm genuinely excited to see how developers leverage Glimmer for complex, multi-step tasks. The focus on reliable tool use is exactly what we need to move beyond simple prompt-response loops. This could really accelerate local AI agent development.