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

Details

Meta's Muse Glimmer is a dense 30B parameter model designed for local and privacy-aware agentic use cases including coding, document analysis, and personal assistants. Its architecture combines a 2B ViT-style Perception Encoder for vision with a 28B text decoder. It is distilled from the larger Muse model and released under Apache 2.0, making commercial deployment straightforward. On agentic benchmarks, Muse Glimmer leads its weight class: MCP Atlas (75.5 vs Gemma4-31B's 54.2 and Qwen3.6-27B's 62.5), GAIA2 (43.3 vs 36.4 and 40.0), and SWE-Bench Verified (76.0 vs 66.6, narrowly behind Qwen's 77.2). Reasoning scores include 94.7 on AIME 2026 and 83.5 on GPQA Diamond. Safety benchmarks show a 26.4% memory violation rate and 28.4% agent attack success rate — intermediate among compared models, warranting review before production agentic deployment. Day-0 support ships in Transformers, llama.cpp, vLLM, and Hugging Face Inference Endpoints. An optional speculative decoding drafter built on DFlash provides faster inference at the cost of additional memory.