Muse Glimmer, a free AI model you can actually run

Meta released a new open-weight model, and this one is different in the way that matters most to me: you can actually run it. Muse Glimmer is 30 billion parameters, dense, multimodal, and released under Apache 2.0, a real open-source license with no user caps and no acceptable-use annex. I read Zuckerberg's announcement, went through Meta's blog post and the independent benchmarks, and then pulled it onto my own machine the same day. Here's the honest version of the story.

What Meta shipped

Zuckerberg announced it on X: a 30B dense model that runs locally, with the weights up on Hugging Face the same day, and credit given to Alexandr Wang and the Meta Superintelligence Labs team. Meta says weights for Muse Spark 1.2, their frontier model, are coming too. Whatever you think of Meta, an American lab shipping true Apache 2.0 open weights again is a step in the right direction, and I say that as someone who doubts Anthropic or OpenAI will follow.

The model itself is tuned for agent work: tool calls, long tasks, failure recovery, multi-step reasoning. It takes text and images. Meta's own use-case list is telling: scheduling, drafting messages, organizing files, following instructions over long contexts. This is not trying to be a frontier chatbot. It's trying to be the reliable little engine inside a local agent.

The real engineering story: 55GB squeezed into 24GB

Unquantized, at full precision, this model needs 55+ gigabytes of memory. That's datacenter territory. Meta shipped it compressed to roughly 4-bit, under 20 gigabytes, and by all appearances they designed the whole thing backwards from one target: fitting the 24GB VRAM class, the RTX 3090s and 4090s and M-series Macs that enthusiasts actually own, while losing very little quality in the squeeze. The compression is the headline. On Ollama it's about a 19GB download.

The benchmarks, honestly

Meta's own comparison table puts Glimmer ahead of Gemma 4 31B and Qwen 3.6 27B on several agentic benchmarks. To their credit, the same page also shows where it loses, including a hallucination rate around 82%, notably worse than Qwen and Gemma. That admission changes how you should use the thing: this is a model you point at tools and files that can tell it when it's wrong, not one you quiz on facts.

On the independent side, Artificial Analysis ran it through their own harness on launch day and placed it mid-pack for its size class on their intelligence index, above Gemma 4 31B, below Qwen 3.6 27B. It's very early, launch-day numbers move as the ecosystem settles, and I trust neither vendor tables nor day-one leaderboards completely. But the shape is consistent: not the smartest model in its class, tuned instead to be steady inside an agent loop. For local AI, that's honestly a good trade.

I ran it the day it dropped

This is the part no press release covers. To get it running I had to update both Ollama and Open WebUI, because support for the model shipped in Ollama the same day as the model itself. That's how day-one this is.

It works. It's also slow on my machine right now: I'm on an AMD Ryzen AI Max box, there's no optimization for that chip yet, and a long writing task took about two minutes. The speed story (Meta's DFlash speculative decoding numbers) currently applies to hardware I don't own. I expect that to change, and when it does I'll re-test. But it's genuinely wild that you can run one pull command, open a browser, and have a brand-new 30B model working in minutes. That used to take hours of fighting.

Who should care

If you're running local AI today, Qwen 3.6 and Gemma 4 are more mature and better optimized, and most people should probably stay on them for now. Glimmer is the one to watch if your use case is agents: tool calling, instruction following, long-context memory, on hardware you own, with your data staying home. That privacy trade is why most of my own daily AI (organizing, drafting, management) runs locally, and it does not need frontier-level output.

And this might all be old news within days: Qwen 3.8 is expected imminently, it's one of the most anticipated local models of the year, and when it lands I'll run the direct comparison. That's the next video: Muse Glimmer side by side against Qwen and Gemma on the same one-shot tests, on the same machine.

Sources

I am documenting a life of interest and curiosity, tech, money, health, and family included. New video daily.