Muse Glimmer 30B

The call

The most consequential local model of the year so far, and the first fully open release Meta has made since it closed the Llama line behind the proprietary Muse Spark in April. Glimmer is distilled from Muse Spark and trained explicitly around the agent loop rather than around chat: plan, call a tool, read the result, keep going, and recover when a call fails. It takes interleaved text and images through a dedicated perception encoder, so an agent can read a screenshot or a chart mid-task. Meta ships it with two 4-bit quantizations and a DFlash speculative-decoding drafter that proposes 16-token blocks in one pass — on their own numbers an RTX 5090 goes from 74.9 tok/s to 233.4 tok/s with the drafter attached. The licence is the other half of the story: Apache 2.0 with no user-count clause, which is more permissive than any Llama release ever was.

When not to use: General chat or knowledge work — it is tuned for autonomous task execution, and its knowledge cutoff is January 4, 2026. Anything under 24 GB: the compression target is a 24 GB envelope, and below that you are pushing the KV cache or the vision encoder off the card. Treat the benchmark table as vendor-run until independent numbers land; Meta's own card is refreshingly honest that Glimmer is roughly in line with Gemma 4 31B and Qwen3.6-27B in its size class rather than ahead of it, and it loses several chem/bio rows to Gemma 4 outright. And note what local execution does *not* fix: keeping data on-device does nothing about prompt injection or an agent taking an irreversible action, which is exactly the risk surface an agentic model raises. Meta recommends human-in-the-loop confirmation on destructive steps and so do we.