Meta releases Muse Glimmer-30B open weights for local agents

Meta's Superintelligence Lab has posted a GGUF build of Muse Glimmer-30B on Hugging Face, a roughly 29.6-billion-parameter multimodal model distilled from a larger Muse Spark and released under Apache 2.0. The interesting number is not the parameter count, it's the packaging: one of the two shipped quantizations targets a 24GB VRAM envelope and, per Meta's own measurements, gives up only about 1.0% of full-precision quality.

The model card lists a run of benchmark wins over comparable open-weight rivals. On MCP Atlas, an agentic evaluation, Meta reports 75.5 and claims the model outperforms Gemma4-31B and Qwen3.6-27B. Coding scores include 51.2 on SWE-Bench Pro and 76.0 on SWE-Bench Verified. Reasoning scores include 94.7 on AIME 2026 and 83.5 on GPQA Diamond. A companion DFlash drafter for speculative decoding is bundled alongside the weights, and Meta says it pushes an RTX 5090 from 74.9 to 233.4 tokens per second, a 3.1x speedup.

For anyone actually building local agents, the practical piece is scaffold compatibility. The card names OpenClaw and Hermes Agent as tested orchestration patterns, alongside deployment through llama.cpp, vLLM, Ollama, LM Studio, and Jan. Image input is handled by a roughly 1.8B ViT-G/14 perception encoder shipped as a separate mmproj file, though the model is text-out only and does not treat audio or video as primary modalities.

There are real caveats. Every benchmark number here is self-reported, without third-party replication yet, and the safety report Meta links to describes the parent Muse Spark model rather than this distilled variant. The card also stops short of explaining what "distilled from Muse Spark" involved in practice, meaning the teacher size and training recipe are not disclosed.

If even the coding and agentic numbers hold up under independent testing, the shape of this release is the story: an Apache-licensed, agent-capable 30B model that fits on a single consumer card is the kind of drop that quietly changes what a small team can ship without renting a hosted frontier API.