If you've been paying $20 a month for ChatGPT — or skipping AI tools entirely because you're not comfortable with your questions floating off to a server farm somewhere — Meta shipped something last week worth knowing about.

Meta released Muse Glimmer on August 10, a 30-billion-parameter AI model. "Parameter" is the technical word for how complex and capable the model is — a higher number generally means more capable. The model is big enough to be genuinely useful on real tasks, small enough to fit on a modern personal computer. It's licensed under Apache 2.0, which means you download it, run it, modify it, use it commercially, and never owe Meta a dollar. No account required. No subscription.

And once it's running on your machine, nothing you type goes anywhere. That's the part worth sitting with.

What "local AI" actually means

Here's the object lesson that might make this click.

When you ask ChatGPT about your salary negotiation, your medical symptoms, or the difficult email you're about to send your landlord — that question is processed on OpenAI's servers. It's in a log somewhere. Their privacy policy has a lot of words about what happens to it. Whether you trust those words or not, the data left your device.

Muse Glimmer processes your question on your own hardware. The only one who knows you asked it is you. It's the difference between writing your diary at home versus emailing your diary to a company and hoping they're careful with it.

Compressed, the model weights fit under 20 gigabytes. It runs on a single consumer GPU with 24-32 gigabytes of video memory — that's roughly an RTX 4090, an RTX 4080 Super, or an Apple M3 Max with 40-core GPU. Supports more than 100 languages. The context window — how much text it can hold in its head at one time — is 128,000 tokens, or roughly 90,000 words. A full novel, more or less.

Where it's strong and where it trails

The benchmarks are mixed, which is the honest answer. According to Artificial Analysis's independent testing, Muse Glimmer scored 94.7 on AIME 2026 (a competition math benchmark) — ahead of comparable open models from Qwen at 94.1 and Google's Gemma at 89.2. On MCP Atlas, which measures how well a model sequences software tools to complete agentic tasks, it scored 75.5 against Qwen's 54.2 and Gemma's 62.5. These are real leads. The model was designed for multi-step agentic work, and the agentic benchmarks reflect that.

Where it trails: Terminal-Bench 2.1 (command-line task completion) came in at 51.7 versus Qwen's 60.7. Graduate-level science questions (GPQA Diamond) scored 83.5 against Gemma's 85.7. OSWorld-Verified — computer-use tasks — scored 65.9 against Qwen's 75.6.

Or maybe not trailing at all, depending on what you're using it for. A 30-billion-parameter model running on consumer hardware is not going to beat frontier cloud models at everything. The more interesting comparison is: can it do the things I actually need, offline, for free?

One real limitation worth knowing: its knowledge cutoff is January 4, 2026. Ask it what happened last month and it'll either say it doesn't know, or — and this is the less ideal outcome — it'll confidently make something up. That's not a Muse Glimmer problem specifically; all AI models do it. But offline models can't be updated in real time the way cloud ones are. For current events, recent prices, or anything that changed this year, you'll need to tell it what you know.

Source spread

• Meta AI Research — Official announcement — hype. Meta's framing emphasizes privacy, offline capability, and agentic use cases. Zuckerberg's "personal AI" thesis runs through it.
• Artificial Analysis — Muse Glimmer benchmarks — builder. Independent benchmark comparisons against Qwen3.6-27B and Gemma4-31B across agentic and reasoning tasks.
• TechCrunch — Zuckerberg's personal intelligence vision — skeptic. Broader context on what Meta's open-source strategy is actually in service of and who benefits most from the move.
• Engadget — Meta's open source Muse Glimmer — builder. Covers hardware requirements, quantization details, and practical setup path.

What's real and what deserves a side-eye

What's real:

• A capable open-weight AI running offline on consumer hardware is genuinely new. This quality of local model wasn't viable a year ago.
• Apache 2.0 licensing is as open as it gets. You can build real products on this without negotiating with anyone.
• The privacy benefit is real, not a sales pitch. For medical questions, legal documents, financial planning, client files — anything you'd rather not have on someone else's server — a local model keeps that data on your device. Full stop.
• Meta's open-source commitment here is meaningful. The weights are public. The code is open. That's hard to take back.

What deserves a side-eye:

• "Consumer GPU" is doing a lot of work in Meta's marketing. An RTX 4090 costs over $1,500. Apple M3 Max systems start above $2,000. If you don't already own the hardware, "free AI" comes with a sizable hardware prerequisite.
• Knowledge cutoff of January 4, 2026 is a real constraint. Seven months of news, product launches, and market changes are not in this model.
• Muse Glimmer was designed for coding and agentic tasks. You can use it as a general chatbot — and it's good — but that's not what it's optimized for.

What to do about it

• If you have a capable GPU or Apple Silicon Mac: Download Muse Glimmer from Meta's Hugging Face repository and run it through Ollama or LM Studio — both give you a clean chat interface without writing any code. Start with a document you'd never want to upload anywhere and ask questions about it. That's the right first test.
• For genuinely sensitive work: Local AI is worth serious consideration if you handle medical records, legal documents, financial plans, or anything client-confidential. The hardware investment pays differently when the alternative is putting client data in a cloud system.
• If you don't have the hardware yet: Bookmark this. Hardware costs are coming down fast. The 30B model of today will run more easily on hardware that costs half as much in two years. The use case is real even if the entry cost isn't there yet.
• Know the knowledge cutoff: Ask Muse Glimmer about anything after January 2026 and treat its answers as uncertain. For current information on any topic, either supplement with a search tool or use a cloud model that has current knowledge.

Further reading

• Meta AI Research — Introducing Muse Glimmer — official announcement with full technical specs
• Artificial Analysis — Muse Glimmer benchmarks — independent benchmark comparison against comparable open-weight models
• TechCrunch — Zuckerberg's personal intelligence vision — context on Meta's broader open-source strategy
• NVIDIA Technical Blog — Run Muse Glimmer with agentic workflows — hardware-specific setup guide for NVIDIA GPUs
• AMD — Run Muse Glimmer on Radeon GPUs — AMD hardware setup guide