Meta open-sources Muse Glimmer, a 30B on-device agent model

A 30 billion parameter open model that is aimed squarely at running agents on a single consumer GPU is a more interesting release than the usual leaderboard-topping flagship, because it points at where a lot of the practical open-weights work is heading. That is the pitch behind Muse Glimmer, which Meta published on its Hugging Face org on August 10, 2026 and wrote up in a companion Hugging Face blog post.

The model is distilled from a larger Muse system down to 30B parameters, released under Apache 2.0, and shipped with vision through a 2B ViT-style perception encoder feeding a 28B text decoder. The collection ships the BF16 weights, a Q4_K_M GGUF quant, an ExecuTorch build for on-device execution, and a 3B assistant model, plus what the blog calls a DFlash block-diffusion drafter for speculative decoding. According to a summary at Open Source For You, the 4-bit compression is what lets the whole stack (weights, KV cache, perception encoder, drafter) run inside a 24 GB or 32 GB consumer GPU envelope rather than a data-center card.

Meta is claiming category wins on a long list of benchmarks: 75.5 on MCP Atlas, 51.2 on SWE-Bench Pro, 94.7 on AIME 2026, 77.0 on IFBench, and 78.8 on Charxiv Reasoning. Those are Meta's own reported numbers on its own selection of benchmarks, so treat them the way you would any vendor scorecard. The useful question is how the model actually performs on your own agent task, not how it stacks up in a curated table.

A couple of gaps in the write-up are worth flagging. The blog states a 32,768 token context length, which is short for the kind of long-horizon agent runs the model is being pitched at. The post also does not explain how the reported safety scores (26.4 on CI Memories Violation, 28.4 on Siren AgentDojo attack success) translate into real-world resistance for an agent that browses, opens files and calls tools on a user's machine.

If a 30B model can genuinely do multi-step tool use, screenshot reading and code editing on one card, a lot of the agent stack that used to live inside paid APIs (the calls, the log data, the per-token bill) moves onto the user's own hardware. That is a real shift for privacy-sensitive teams, students, and anyone in a region where the hosted equivalents are expensive or restricted.