Meta is back with Muse Glimmer: local, agentic, multimodal, and open source

Source: HuggingFace Date: 2026-08-10 URL: https://huggingface.co/blog/muse-glimmer

Summary

Meta released Muse Glimmer, a 30B-parameter (2B vision encoder + 28B text decoder) multimodal model distilled from a larger Muse model, under Apache 2.0, targeting local/on-device agentic deployment. It reports 51.2% on SWE-Bench Pro and 75.5 on the MCP Atlas agentic benchmark, ahead of comparably-sized Gemma4-31B and Qwen3.6-27B, with day-one support in transformers, llama.cpp, vLLM, and HF Inference Endpoints.

Implications

Open-weight capability clock. A frontier lab shipping a sub-35B, Apache-licensed, agent-and-coding-competitive model that beats larger open peers on SWE-Bench Pro and an MCP-native benchmark is a real clock tick — smaller, permissively-licensed, and explicitly tuned for tool use and local execution rather than raw scale. Worth pairing against DeepSeek V4 and Kimi K2.6 in the open-weight ladder.

Agentic engineering patterns. Direct MCP Atlas benchmarking (rather than only chat/coding evals) signals MCP tool-use is becoming a standard axis vendors compete on, and day-one llama.cpp/vLLM support lowers the bar for local agent deployments to adopt it immediately.