Meta has released Muse Glimmer, a 30-billion-parameter open-weights model built for a specific job: running agents locally. Announced on 10 August, it arrives under a permissive Apache 2.0 licence — the terms that let anyone use, modify and ship a model commercially — with the weights available to download on Hugging Face and support already wired into the tools people run models with locally, including Ollama, LM Studio and vLLM.

The pitch is in the positioning. Meta describes Glimmer as "optimized for always-on local agent workflows" — not a general chatbot but a model sized and tuned to sit on your own hardware and do the unglamorous work agents actually need: function calling, coding, and acting as an LLM-as-a-judge to grade other models' output. At 30B it is small enough to run on a well-specced local machine and large enough to be useful — Meta says it fits on a Mac or PC with a single consumer GPU, compressing to under 20 GB at roughly 4-bit — which is the band the local-agent use case has been asking for.

Meta puts Glimmer up against two obvious rivals in that size class — Google's Gemma 4 (31B) and Alibaba's Qwen 3.6 (27B) — and reports competitive results across agentic, coding, multimodal, safety and reasoning benchmarks, naming SWE-Bench, τ-Bench, MCP-Atlas and DeepSearch QA among them. As ever, these are the vendor's own numbers on the vendor's own chosen benchmarks; the independent evaluations that actually settle a model's standing come later, and this desk waits for them before calling any leaderboard. What is not in dispute is the licence and the availability: Apache 2.0 and a Hugging Face download are facts, not claims.

That combination is the story. Open weights have been the year's quiet through-line, and the front of that race has moved from "can we match the closed labs on raw capability" to "can we give people a capable model built for a concrete job they can run themselves." Glimmer is the second kind: not a bid for the top of a general-intelligence leaderboard, but a tool aimed at the developer who wants an agent that runs on their own box, under their own control, without a per-token bill or a data round-trip to someone else's cloud.

It also lands on a pointed day. This morning's Daily Update covered coding agents being handed more autonomy by default; a 30B open model you can run locally is the other side of that coin — the infrastructure that lets you keep an autonomous agent on hardware you own rather than renting it. Meta says Glimmer was assessed under its Advanced AI Scaling Framework, the company's internal safety-evaluation standard; what that assessment covered in practice, and how the model behaves once the community has it in hand, is the part worth watching now that anyone can pull the weights.

The open-weights field does not lack for 30B models. What it has lacked is one that treats "run agents locally" as the design brief rather than an afterthought. Whether Glimmer delivers on that brief is a question the independent numbers — and the developers now downloading it from Hugging Face — will answer over the coming days.