The AI That Runs on Your Own Machine Is Now Good Enough to Matter
Meta's new open model runs on a single consumer GPU, handles complex agent tasks, and costs nothing per query — a real shift for any business that cares about privacy or API bills.
- 30B
- parameters in the model
- 24 GB
- GPU memory needed to run it
- 75.5
- agent benchmark score (MCP Atlas)
- Apache 2.0
- license — free for commercial use
Until last week, running a genuinely capable AI agent meant sending your data to someone else's server and paying for every query. Meta changed that on August 10 when it released Muse Glimmer — a capable AI that runs entirely on your own hardware. It is free to use commercially under the Apache 2.0 license. No monthly fees, no per-token charges, no usage caps.
Muse Glimmer has 30 billion parameters — a measure of how much it has learned. That is large enough to handle real work: calling tools, reading long documents, writing and running code, and judging its own outputs. On the main agent benchmark (MCP Atlas), it scores 75.5, against 54.2 for Google's Gemma4-31B and 62.5 for Qwen3.6-27B. It beats similar-sized rivals at the work agents actually do. It fits on a 24 GB graphics card — the kind found in a mid-range gaming PC or a recent MacBook Pro.
For businesses in law, finance, or health, the privacy angle is the biggest deal. Every document you send to a cloud AI is data you hand to another company's server. With Muse Glimmer, your data never has to leave your building. A law firm can run contract-review agents, a fintech team can check compliance logic, a clinic can analyse patient notes — all on a machine they own and control.
The cost model changes too. Cloud APIs charge by the token — the small chunks of text the model reads and writes. An agent that retries, loops, or runs hundreds of checks a day can rack up a surprising bill. With a local model, the cost becomes fixed: buy the hardware once, run it as much as you need. If your agent volume is growing or your data is sensitive, Muse Glimmer is worth a test run this week — the weights are free to download from Hugging Face.