Meta put a 30-billion-parameter agent on a single graphics card. I unplugged the internet and gave it my to-do list.

Meta dropped Muse Glimmer on a Monday, which is the only part of this that went to plan. The rest of my week reorganized itself around a single question that the release is built to provoke: what happens when a capable agent stops living in someone else's data center and starts living on the machine on your desk. I downloaded the weights the night they landed, pulled the ethernet cable out of my desktop like a man staging his own alibi, and spent the next day and a half handing a 30-billion-parameter model the kind of small, annoying jobs I would normally farm out to a cloud API. This is not the full week I usually give a tool before I write about it. The model is barely older than this article. But some things you learn in the first hour, and the first thing I learned is that offline is not a feature. It is the whole argument.

Here is what Meta actually shipped, stripped of the launch-post adjectives. Muse Glimmer is a 30-billion-parameter open-weights model released under Apache 2.0, which is the permissive license that lets you use it commercially without asking anyone. Meta's Superintelligence Labs says it distilled the thing down from a larger model it calls Muse Spark, then quantized it to roughly four-bit precision so the memory footprint drops under 20GB. The practical translation: a build that fits inside the 24GB to 32GB of video memory on a single high-end consumer graphics card, or the unified memory of a recent Mac. It takes text and images, speaks a claimed hundred-plus languages, and it is tuned specifically for the unglamorous mechanics of being an agent: calling tools, following a function schema, running a multi-step task, and recovering when a step fails. That last item is the one I came to test, because failure recovery is where every agent I have ever used goes to die.

Getting it running is easier than it should be

I expected the setup to be the punishing part. It wasn't. Meta shipped Glimmer into the ecosystem that already exists rather than inventing a new one, which is the single most respectful thing a company can do to the person who has to install its software. The quantized builds are on Hugging Face, and they load in the local runners people already have open: Ollama, LM Studio, llama.cpp for the command-line crowd, MLX and ExecuTorch if you are on Apple silicon. I ran the K-Quant build, which lands around 17GB, and it sat comfortably inside a 24GB card with room left over for the context. From cold download to the model answering its first question was about twenty minutes, most of which was the download.

The speed surprised me more than the size. Meta bundles a speculative-decoding trick it calls DFlash, and it claims something like a 3x generation speedup on the newest desktop cards and a 1.5x to 1.8x bump on Mac laptops. I can't verify the multiplier, but I can tell you the felt experience: on a good desktop GPU it generates faster than I read, which is the threshold past which a local model stops feeling like a science project and starts feeling like a tool. On a laptop it is slower, closer to watching someone type, but it never stalled. For a model running entirely on hardware I already owned, with the internet physically disconnected, that is a genuinely strange thing to sit in front of.

The Tuesday test: three real jobs, one bad afternoon

A benchmark tells you how a model does on a test written to be graded. I care about the Tuesday. So I gave Glimmer three jobs I actually needed done, wired it up to real tools, and watched where the seams showed.

The first job was code. I pointed it at a small, real repository of mine with a real, boring bug: a date-parsing function that fell over on one timezone and quietly returned the wrong day. This is exactly the sort of thing the cloud agents are sold on. Glimmer read the file, found the function, reasoned out loud about the edge case, and proposed a fix that was correct. Not almost correct. Correct, on the first pass, with a one-line explanation that was actually about the bug and not about how helpful it was being. For a model I was running for free on my own silicon, that was the moment the exercise stopped being a novelty.

The second job was tool use, which is the part Meta is really selling. I gave it a set of local functions — read a directory, move a file, write a short summary — and a plain instruction: sort the mess in my downloads folder into sensible subfolders and leave me a note about what it did. This is a multi-step task with real consequences, the kind where a wrong guess doesn't produce a bad sentence, it produces a lost file. Glimmer called the functions in a sane order, respected the schema I gave it, and got about eighty percent of the way through cleanly. Then it hit a file with an ambiguous name, and instead of guessing, it stopped and asked. I want to be precise about how rare that is. Most agents, faced with ambiguity, invent a confident answer and move on, and a confident wrong move wearing a tidy folder structure is just a mess you now have to reverse-engineer. Glimmer noticed the ambiguity and handed it back. That is the failure recovery Meta advertised, and it is the difference between a tool you can leave alone and one you have to babysit.

A confident wrong move wearing a tidy folder structure is just a mess you now have to reverse-engineer. Glimmer noticed the ambiguity and handed it back.

And then the third job went sideways, which is why I am glad I ran three. I asked it to do something genuinely multi-step: read a long PDF, pull out the five dates that mattered, and draft calendar entries for each with a one-line context note. It read the document fine. It found the dates fine. But somewhere in the handoff between 'summarize this' and 'now turn each item into a structured calendar entry,' it lost the thread — it started well, drifted into re-summarizing the document it had already summarized, and produced three good entries and two that were confidently attached to the wrong context. This is the seam I have watched every agent fray at: not the single clean task, but the moment two tasks touch and the model has to hold the output of one while doing the other. Glimmer is better at this than most 30B models have any right to be. It is not as good as the big cloud models, and the gap shows up exactly here, in the connective tissue between steps.

What 'four-bit' actually costs you

There is a temptation, when a model this small does something this good, to declare that quantization is free and the frontier is now a download away. It isn't, and pretending otherwise does the reader a disservice. Compressing a model to four bits is a lossy act, and Meta's own framing — 'minimal to no degradation on agentic tasks' — is doing quiet work in that word 'minimal.' The damage from quantization doesn't spread evenly; it pools in the hard cases, the long-context reasoning and the rare edge that you bought a 30B model to handle in the first place. In my day and a half, the degradation showed up not as obvious wrongness but as a slightly shorter attention span. Glimmer is sharp on the task in front of it and gets vaguer the more you ask it to hold at once. That is a completely reasonable trade for a model that runs on your laptop. It is not a free one, and the launch language rounds it down to zero.

Meta also says Glimmer beats two specific rivals — Google's Gemma 4 at 31B and Alibaba's Qwen 3.6 at 27B — across the agentic, coding and reasoning benchmarks it chose to publish. Take that the way you take every first-party benchmark: as the home team's box score, on a field the home team lined. The comparison being to exactly two models, both in the same weight class, tells you the framing is 'best small open model for local agents,' not 'competitive with the cloud.' That is a real and useful claim. It is also a narrower one than the headline energy suggests.

The part that isn't about the model

Spend a day with Glimmer running offline and you start to feel the actual pitch, which is not really about capability at all. It is about where the work happens. Every task I ran cost me nothing per token, because there was no meter. Nothing I fed it — my code, my files, that PDF — left the machine, because there was no network. There is no rate limit, no model-deprecation email, no morning where the thing you built your workflow on gets quietly retired by a company optimizing its serving costs. For a whole category of work — private documents, offline environments, anything you'd rather not narrate to a third party's logs — that changes the math in a way no benchmark captures.

That is also the honest boundary of who this is for. The always-on private agent is a real thing now, but it asks for a real graphics card — a recent high-end GPU with 24GB or more, or a well-specced Mac — and it rewards people who are comfortable wiring up their own tools. If you have that hardware and that inclination, Glimmer is the most convincing local agent I have run, and it is free. If you are waiting for the phone-in-your-pocket version, this isn't it yet; the memory math still wants a desktop. And if your work genuinely needs the frontier — the long, tangled, many-step reasoning where the cloud models still pull ahead — you will feel the ceiling in an afternoon, right at the seam where two tasks meet.

The verdict

I keep a running list of tools I install and cancel, and the test I apply is simple: a week later, is it still in the workflow or is it in the graveyard. Glimmer is too new for me to honestly file it either way — ask me in a week. But I can tell you it survived first contact better than anything in its size class I have run, and that it did so with the internet unplugged, which a day ago I would have called a demo and today I am calling a Tuesday. The model is not going to make you cancel your cloud subscription. What it does is quieter and, for the right person, more interesting: it makes a genuinely useful agent something you can own outright, run in a room with no signal, and never explain to anyone. Meta gave that away for free under a license that lets you build on it. The catch is the graphics card and the ceiling, and both are exactly where you'd expect them to be. For once, the seam is in the honest place.