--- title: "The agent can live in the building now" date: 2026-08-12 url: https://remiam.co.uk/notes/the-agent-can-live-in-the-building tags: [AI, Agents, Infrastructure] read_time_minutes: 6 description: "Meta released Muse Glimmer, a 30B open-weights agentic model that fits a 24GB consumer GPU. What an offline agent actually changes for the systems we install on site." --- # The agent can live in the building now *Published 2026-08-12 · 6 min read · by Liam (Remiam)* Meta put a 30 billion parameter agentic model out under Apache 2.0 on 10 August, small enough for one consumer GPU. The licence is the news, not the benchmark table. On 10 August Meta released Muse Glimmer, a 30 billion parameter model under an Apache 2.0 licence, aimed squarely at always-on local agent workflows. Quantised it drops under 20GB, which puts it inside a 24GB or 32GB consumer graphics card, or a Mac. It is distilled from the much larger Muse Spark using logit distillation, and it ships with integrations for llama.cpp, MLX and ExecuTorch alongside Ollama, vLLM and SGLang (Meta AI Research, 2026). The benchmark table is the part everyone will argue about, and it is the least interesting thing here. Apache 2.0 plus 24GB is the sentence that matters. A model you can use commercially, modify, and redistribute, running on a machine you can carry into a building, is a different category of component from a model you rent through an endpoint. ## What changes when the weights are yours Three things, and none of them are about intelligence. There is no per-token bill, so the cost of a feature stops scaling with how much people use it. There is no egress, so the data never leaves the room, which turns a six week legal conversation into a diagram. And nobody can deprecate the model underneath a system you installed two years ago, because you hold the weights. That last one is the quiet risk in every AI feature shipped in the last three years. We have all built against an endpoint that was retired, renamed, or repriced. On a marketing site that is an afternoon. On a system running in a venue with a support contract, it is a site visit and an awkward email. ## The installs this actually unlocks Most of our work sits in places with bad connectivity and firm rules about data. A trade stand on a hotel network. A retail floor where the guest wifi is saturated by half past ten. A clinic where nothing that identifies a patient is allowed to leave the building. A visitor attraction in a listed building with one flaky uplink and no appetite for a second. Until now the honest advice in those rooms was to keep the clever part small and the network path short, because an agent that stops working when the line drops is worse than no agent at all. A 30B model on a box in the comms cupboard changes the answer. The feature degrades to slower, not to absent. > The question was never how clever the model is. It was what happens to the install at ten past nine when the line goes down. ## What we would not do with it Two cautions. First, this is a distilled model, compressed further to fit the card. It is tuned for tool use, long tasks and recovering from its own mistakes, not for being the smartest thing in the room. Ask it to be a general oracle and you will be disappointed in a way that is nobody's fault but the specification. Second, a local box is a box. Someone owns it, patches it, cools it, and is on call for it. The model just became the cheap part of that sentence. Operations did not. If you are pricing an offline agent purely on the saving from the API bill, you have costed half of it. ## What we would spec - One machine, one job. Do not put the agent on the box that runs the till, the display wall or the door. - A small, explicit tool surface. Every tool it can call is a thing it can get wrong at three in the morning. - Log every tool call with a timestamp and an outcome. The point of a local system is that you can actually inspect it. - A hosted fallback for the hardest few percent of requests, with a hard rule about what data may take that path. - Treat a model update as a deployment, with a version, a test pass and a rollback. It is not a subscription any more. If you are scoping an AI feature for a physical space this year, price the offline version alongside the hosted one. Two weeks ago that was a research project. It is now a purchasing decision, and for a lot of the buildings we work in it is the better one. ## References 1. [Meta AI Research, Introducing Muse Glimmer, 2026](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) 2. [Meta, Muse Glimmer model page, 2026](https://developer.meta.com/ai/models/muse-glimmer/) 3. [VentureBeat, Meta returns to open source with Muse Glimmer, 2026](https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now)