The Biggest Free AI Model Ever. Read the License First

Moonshot AI has published the weights for a 2.8 trillion parameter model on Hugging Face, free for anyone to download. It is the largest open-weight AI model released to date, and on the company’s own benchmark table it trades blows with the newest systems from OpenAI and Anthropic.

If you run a small business, the useful reaction is not excitement and it is not alarm. You will almost certainly never download this model, and that is fine, because the release still moves what you pay and who you are able to switch to. Two words dominate the coverage, and both mislead. “Free” does not mean what most owners assume, and “Chinese” does not mean what most owners fear. What open source AI for small business actually delivers is leverage, and it arrives through your existing vendors rather than through your own servers.

What Moonshot actually released

Kimi K3 is a mixture-of-experts model, which means that despite the 2.8 trillion parameter headline it only activates 16 of its 896 experts on any given token, according to the official model card. That design is why a model this large can be served at a sane cost. It carries a one million token context window and 93 layers.

On the benchmarks Moonshot published, K3 scores 93.5 on GPQA Diamond against 91.0 for Claude Opus 4.8 and 93.5 for GPT-5.5, 67.5 on DeepSWE against 59.0 and 67.0, and 91.2 on BrowseComp against 84.3 and 84.4 (model card). Read those numbers with the appropriate squint. They are self-reported by the company selling the model, measured on tests the company chose. Independent evaluation takes weeks. The honest summary is that K3 is credibly in the same class as the frontier US models, not that it has beaten them.

Downloading it is not a realistic option for a small business regardless. The full weights run to hundreds of gigabytes and serving them requires a multi-node cluster of high-end GPUs. This is not a thing you install on the office computer, and any coverage implying otherwise is selling you something.

What open source AI for small business actually gets you

The interesting document is not the model card. It is the license, and it is unusually kind to exactly the people reading this.

The commercial restrictions in the Kimi K3 License do not apply to internal use, which the license defines as any use that does not make the model, its outputs, or its underlying capabilities available to third parties. Using it to draft your quotes, summarise your inbox, or analyse your sales data sits squarely inside that exemption.

The restrictions land on the resellers instead. A company running a model-as-a-service business has to negotiate a separate agreement with Moonshot once its revenue passes 20 million US dollars over any consecutive twelve months. Products above 100 million monthly active users, or above 20 million dollars in monthly revenue, have to display “Kimi K3” prominently in their interface. Those are thresholds written for platforms, not for a fourteen-person contractor.

So the practical answer to “is it free for my business” is yes, with a caveat that has nothing to do with licensing and everything to do with logistics: you will consume it through somebody else’s hosting, and that host sets your actual price.

Will my data go to China?

This is the question owners actually ask, and it deserves a straight answer rather than reassurance.

Where a model was trained and where it runs are two separate facts. Open weights are what separates them. Because the weights are published, any provider anywhere can load K3 onto their own hardware and serve it, which means a US or EU host can run this model on servers in your jurisdiction, under their contract and their data policy. Your prompts go to whoever is hosting, not to Moonshot.

The reverse is also true and worth stating plainly. If you use Moonshot’s own apps or its paid API, you are sending data to Moonshot under Moonshot’s terms, and that is a different decision with different implications. The license itself notes that use through Moonshot’s official products and certified inference partners sits outside the commercial restrictions, which tells you those are distinct routes.

The question to ask a vendor is therefore not “which model do you use” but “which model, hosted by whom, on servers where, and can I see it in writing.” That question works for every AI tool you buy, not just this one. If you have not yet written down what staff may and may not put into an AI tool, that gap matters more than any model choice, and we covered how far adoption has outrun the rules earlier this month.

Why a free Chinese model makes your American software cheaper

Here is the second-order effect, and it is the one that reaches your invoice.

When a frontier-class model becomes downloadable, it stops being a moat and starts being a component. Any hosting provider can serve it, and they compete with each other on price and latency rather than on exclusive access. That competition drags down the price of the intelligence you rent, including from vendors who never touch K3, because they now have to justify their premium against a free alternative that scores in the same range.

We wrote last week about the model price war you probably missed, where three frontier launches in four days produced no price increases. This is the same pressure arriving from a different direction. That was vendors declining to raise prices. This is the floor underneath those prices dropping.

The strategic read for a small business is about optionality rather than savings. Treat the model as a swappable part. Before you sign anything annual, ask whether your vendor can switch the underlying model without breaking your workflows, and whether you can export your prompts, your data, and your configurations if you leave. A vendor who cannot answer that is selling you lock-in at a moment when the thing they are locking you into is getting commoditised.

What to actually do this month

Nothing urgent, which is itself the honest answer. There is no button to press today.

Three things are worth doing anyway. Ask your current AI vendors which models they can serve and where those run, and get it in writing. Check whether any annual renewal is coming up in the next quarter, and delay committing if the contract does not let you move. And if you have been waiting for AI to get cheap enough to justify a real experiment, note that the cost curve is still bending downward, which means the argument for waiting gets weaker every quarter, not stronger. Our practical guide to getting real results covers where to start if that is where you are.

The larger pattern is the one worth sitting with. A five-person shop and a five-thousand-person company can now rent close to the same intelligence, and the gap between them keeps narrowing. That gap used to be the whole advantage of being big. The advantage that remains is knowing what to point it at, and that has never been a question of budget.

Frequently Asked Questions

Can I actually use Kimi K3 for free in my business?

Yes for internal use. The Kimi K3 License exempts internal use, defined as any use that does not make the model, its outputs, or its capabilities available to third parties, from its commercial restrictions. You will still pay a hosting provider to run it, because the model is far too large to run on ordinary business hardware, so “free” refers to the licence rather than to your bill.

Is Kimi K3 better than ChatGPT or Claude?

On Moonshot’s own published benchmarks it scores comparably to GPT-5.5 and above Claude Opus 4.8 on several tests, but those numbers are self-reported by the company that built it and have not yet been independently verified. The reasonable conclusion is that K3 is in the same performance class as the leading US models, which is itself a significant change, rather than that it is definitively better.

Does using a Chinese AI model mean my data goes to China?

Not necessarily, because where a model is trained and where it runs are separate. Since the weights are public, providers in the US or EU can host K3 on their own servers under their own data policies, in which case your data goes to that host. If you use Moonshot’s own apps or API directly, then you are sending data to Moonshot under its terms, which is a different decision.

Should I switch my AI tools because of this release?

No, not on the strength of a model release. The useful response is to check that your existing vendors can swap models underneath you and that you can export your data if you leave. The value of this news to a small business is negotiating leverage and falling prices over the coming months, not an immediate migration.

We keep hearing that the gap between small and large companies is closing because everyone can rent the same intelligence. If that is true, what is the advantage you still have that no amount of compute can buy?