An AI model that runs entirely on your own machine. No internet needed, no subscription, no cloud quietly logging what you ask. Sounds like wishful thinking – but it's been real for a few weeks now.
Meta just releasedMuse Glimmer, a model built to do exactly that. 30 billion parameters, fully open under the Apache 2.0 license, designed for so-called "agentic" tasks: getting things done, calling tools, holding up through long task chains — not just friendly chit-chat.
What makes Muse Glimmer different?
Most big AI models run in massive data centers because they're simply too large for normal hardware. Muse Glimmer is deliberately built not to need that. Meta quantized the model (technically: reduced the precision of the numbers inside it without wrecking quality), shrinking it from 55 gigabytes down to 18–20.
That's just the model data itself, though. Add memory for ongoing conversations and you're quickly at 24 gigabytes of graphics memory — Meta cites 24 to 32 gigabytes as the target. Translation: a single graphics card, yes, but one from the top shelf. Not a data center. But not the office PC either.
The context window (how much text the model can "remember") sits above 120,000 tokens — solid for a local model. It handles images too: those 30 billion parameters include a small vision component, so you can throw it a screenshot or a photo and ask questions about it. And all of it fully offline. No internet, no data sent to Meta, no server logging what you ask.
What's free — and what costs money?
The model itself: completely free. The weights are open on Hugging Face, anyone can download them. No subscription, no API costs, no credit card.
The catch: you need the hardware for it. And more of it than "runs locally" suggests. You won't find those 24 gigabytes of graphics memory in an off-the-shelf laptop — you'll find them in the top price bracket. Used is cheaper, but still no bargain. If you already have Ollama or OpenWebUI running at home, you're ahead of the curve. If you're starting from zero, do the honest math on the purchase price versus a few years of a cloud subscription before you say "free."
Can I try this right now?
Yes — if your hardware plays along. Ollama and LM Studio support Muse Glimmer directly; if you go deeper, there are optimized integrations for llama.cpp, MLX (for Apple machines) and ExecuTorch (for mobile devices). If you've already worked through our Ollama article, you just need to load the right model file — the rest of the pipeline is already in place.
Nvidia measured over 20,000 tokens per second in its own tests. Feel free to forget that number — it comes from the data center, where hundreds of requests are processed at once. On actual home hardware it looks like this: around 84 tokens per second on a current Nvidia graphics card, roughly 24 on an AMD laptop chip, 26 to 50 on current MacBooks. Those are vendor measurements too, taken with the compressed version of the model — your own number depends on the card, the context length, and how heavily the model is compressed. As a ballpark, they still hold up.
For context: a token is roughly a syllable. 24 tokens per second means text appears faster than most people type — but noticeably more sedate than ChatGPT in the cloud. Plenty for reading along. For tasks where the AI grinds away on its own for minutes, it gets tedious.
Worth knowing if you already tinker with AI agents: Muse Glimmer works with OpenClaw, a tool that orchestrates them. We use OpenClaw ourselves — brilliant in principle, but the token counter reliably goes berserk. That particular problem disappears with a local model, because you're no longer paying for tokens, you're paying for electricity. We've been meaning to write about OpenClaw for ages. Whether that article ever gets finished is another matter — the draft has been sitting around for months, looking at us reproachfully.
For me — or just for developers?
Honestly: leans toward tech-savvy folks. If you've never self-hosted a model before, start with Ollama or OpenWebUI first — the entry barrier is lower, and Muse Glimmer ultimately runs through the same tools anyway. But if you already know those steps, Muse Glimmer hands you a current, genuinely capable model for agentic tasks that would overwhelm a plain chat AI.
And yes: this is another chapter in the "why Linux makes AI self-hosting more relaxed" book. If you're already comfortable with Docker, llama.cpp, or Ollama, you know exactly what that means.
The takeaway
Muse Glimmer isn't a replacement for ChatGPT or Claude in daily use — the cloud competition is simply more convenient right now. But as proof that "local, open, powerful" no longer has to be a contradiction, it's a strong signal. The future of AI isn't only happening in data centers — part of it now runs on your own machine, too. If you've got the hardware for it, it's worth a look.