Meta put a working agent on your laptop
Meta dropped Muse Glimmer yesterday, a 30B open-weights model built for one job: running agents locally, all day, on hardware you already own. Apache 2.0. Weights are on Hugging Face right now at meta-models/Muse-Glimmer-30B, GGUF included.
The numbers are the interesting part. MCP-Atlas 75.5, DeepSearch QA 74.6, SWE-Bench Pro 51.2, Gaia2 43.3, IFBench 77.0, AIME 2026 94.7. Those are full-task agentic benchmarks, not multiple choice trivia. A 30B model that finishes end-to-end tasks inside a scaffold, writes and debugs code, recovers from its own failures, and handles multimodal input across 100+ languages. Quantized to 4 bits it fits under 20GB, and it ships with a speculative decoding companion called DFlash that gets 3.1x faster generation on an RTX 5090.
How they got there matters: Glimmer is distilled from Muse Spark, the big teacher, with logit distillation applied across pre-training, mid-training and post-training rather than just a final SFT pass. That is the third paper-shaped result this month arguing the same thing β the bottleneck for small agentic models is signal quality, not parameter count.
The strategic read is harder to miss. Every other lab is selling agent tokens by the million. Meta just gave away a model whose entire pitch is that it never calls a cloud. If always-on personal agents are the next interface, Meta wants that interface running on your machine and nobody's meter. llama.cpp, MLX and ExecuTorch integrations are listed as coming.
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
← Back to all articles
The numbers are the interesting part. MCP-Atlas 75.5, DeepSearch QA 74.6, SWE-Bench Pro 51.2, Gaia2 43.3, IFBench 77.0, AIME 2026 94.7. Those are full-task agentic benchmarks, not multiple choice trivia. A 30B model that finishes end-to-end tasks inside a scaffold, writes and debugs code, recovers from its own failures, and handles multimodal input across 100+ languages. Quantized to 4 bits it fits under 20GB, and it ships with a speculative decoding companion called DFlash that gets 3.1x faster generation on an RTX 5090.
How they got there matters: Glimmer is distilled from Muse Spark, the big teacher, with logit distillation applied across pre-training, mid-training and post-training rather than just a final SFT pass. That is the third paper-shaped result this month arguing the same thing β the bottleneck for small agentic models is signal quality, not parameter count.
The strategic read is harder to miss. Every other lab is selling agent tokens by the million. Meta just gave away a model whose entire pitch is that it never calls a cloud. If always-on personal agents are the next interface, Meta wants that interface running on your machine and nobody's meter. llama.cpp, MLX and ExecuTorch integrations are listed as coming.
https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Comments