Meta's New AI Agent Is Free. Your Hardware Bill Isn't.

For two years, running an AI agent has meant renting somebody else's brain, by the token, on somebody else's server. On 10 August 2026, Meta handed businesses the deed to their own.

What did Meta actually release on 10 August 2026?

Meta Superintelligence Labs shipped Muse Glimmer this morning: a 30-billion-parameter AI model, open-weight, licensed under Apache 2.0, free to download and run commercially with no permission needed from Meta. It reads images as well as text. It holds 128,000 tokens of context. And unlike a general chatbot, it was built specifically for agent work: planning a multi-step task, calling tools, checking its own output, and recovering when a step fails.

The weights are on Hugging Face. Day-zero support landed in Ollama, llama.cpp, vLLM, and Unsloth. Meta says optimized builds for AMD, Arm, Intel, and Nvidia hardware are coming "in the coming days," alongside listings on OpenRouter, Together AI, and Fireworks AI.

Here is the detail that matters more than the spec sheet: Muse Glimmer was distilled from Muse Spark, the same closed model Meta rents out through its paid Model API. Meta named Muse Spark as the teacher model in its own training notes. In plain terms, Meta took the brain businesses already pay to use by the token, compressed it, and gave a copy away.

Why should a small business care about a 30-billion-parameter model?

Because the number that matters for a business is not parameter count. It is what the model does with a tool in its hand.

Meta's own benchmark chart, numbers you should hold loosely because Meta published them itself, puts Muse Glimmer well ahead of the two other open models in its weight class on exactly the tasks an agent needs: reading a tool's output, deciding the next step, and not giving up when something breaks.

To be fair to the competition: Qwen 3.6 pulls back in front on tasks that lean on sustained terminal work and controlling a desktop, and it edges Muse Glimmer on SWE-Bench Verified by about a point, 77.2 to 76.0. A model that is honest about where it loses is a model worth trusting on where it wins. For a bookkeeping practice, a customer-support desk, or a scheduling agent, tool use and structured reasoning are the job. Clicking around a screen is not, yet.

Isn't a "free" AI agent just a marketing headline?

Fair question. Here is the honest answer: this one is not, and the contrast with Meta's own paid product proves it.

Four days before this launch, Meta's paid coding agent, built on the same Muse Spark family, was offering a discounted rate of roughly two cents a query on the condition that Meta could train on your code. That is the standard cloud AI deal: cheaper tokens in exchange for your data.

Muse Glimmer carries none of that. Apache 2.0 has no revenue clause and no list of excluded countries, a real restriction some other "open" model licenses still carry in the fine print. It is the same license Kubernetes ships under. Once you download the weights, there is no meter, no account, and no clause about reading your files. For a bookkeeping practice sitting on client bank statements, payroll runs, and VAT filings, that is not a technicality. It is the difference between a client's financial history sitting on a server you do not control and staying on a machine in your own office.

Can you actually run this today?

This is where the launch-day story gets honest, and where most coverage stopped short. The compressed build is roughly 21GB. Meta's own documentation says it is optimized for 24 to 32GB of consumer memory. A independent test from the channel Hyperautomation Labs put that claim on the exact minimum: a MacBook Pro with 24GB of memory.

The model loaded. Then the machine did not. Free memory fell from 57 percent to zero, the health-check log meant to fire every two seconds started arriving 11, then 30, then 62 seconds apart, and the mouse stopped responding. The tester hard-recovered and deleted the model. His conclusion, which matches the spirit of Meta's own "24 to 32GB" range: 24GB technically loads Muse Glimmer, but 32GB is the practical floor for a machine you still want to use for anything else at the same time.

What about the cloud version Meta promised?

If your business does not own qualifying hardware, the launch post says you can rent Muse Glimmer instead, through OpenRouter, Together AI, Fireworks AI, Meta's own Model API, or Ollama Cloud.

As of the hours after launch, none of them had it live. A live check of OpenRouter's full 400-model catalogue turned up only Muse Glimmer's paid siblings, not Glimmer itself. Together AI and Fireworks AI both returned a bare "not found" on the model page. Meta's own API, checked with a paid key, listed three models, none of them Glimmer. Ollama Cloud's 16-model catalogue had no Glimmer either. Zero routes live out of five, hours after the announcement.

That will change within days, not months, models this well covered rarely stay unhosted for long. But it means the honest answer for a business without a 32GB machine or a serious GPU, today, is: not yet. Plan for it, do not build on it this week.

How to actually put a local AI agent to work, step by step

1. Check your hardware against the table above. If you are on 32GB or more, or a single GPU with 24GB+ of VRAM, you qualify today. If not, this is a "watch and wait" item for your business, not a "do it this afternoon" one.
2. Install Ollama 0.32.7 or later. Earlier versions cannot pull the model; this tripped up even the reviewers testing it on launch day.
3. Pull the model and pick one narrow task first. Do not aim Muse Glimmer at "run my whole back office." Aim it at one job: reading a folder of bank statement PDFs and drafting a month-end reconciliation summary, or triaging a client email inbox into "needs a human today" and "can wait."
4. Wire it into an existing agent harness rather than building your own loop from scratch. Open-source options like OpenClaw and Hermes, and coding-agent tools like Claude Code or Codex, can already point at a local model instead of a cloud API.
5. Keep a human signing off on anything that leaves the building, a reconciliation before it is filed, a reply before it is sent. A local model removes the data-privacy risk of the cloud; it does not remove the need for a review step while you are still learning what it gets wrong.
6. If you don't have the hardware yet, set a calendar reminder for two weeks out and check OpenRouter or Ollama Cloud again. Hosted access is very likely to land fast, and it removes the hardware question entirely.

What you have after this

A week from now, you know whether your current machine even qualifies, and you have run one narrow task on it or found the smaller model that fits instead.

A month from now, one real, recurring piece of your back office, a reconciliation draft, an inbox triage, a first-pass document read, runs on a model you own outright, on a machine that never sends a client's bank statement anywhere.

Six months from now, when the next "free open model" launch shows up in your feed, you will know exactly which question to ask before you believe the headline: not "is it free," but "what does my hardware bill actually say."

Same practice. Same clients. A completely different question at the top of your to-do list.