Kimi K3 is now a free download. It weighs 1.56 terabytes

Moonshot AI, a Chinese company, has released the complete files behind its newest artificial intelligence model. Anyone can download them at no cost, run the model on their own machines, and charge other people for access to it.

The model is called Kimi K3. On independent testing it scores close to the best systems from OpenAI and Anthropic, which cost money to use. This one does not.

The obstacle is size. The files on Moonshot’s repository total 1,560,936,091,448 bytes, or 1.56 terabytes. A machine that can hold all of that at once costs more than most companies in Kenya spend on computing in a year.

What Kimi K3 is

Moonshot AI is a Beijing company founded in March 2023. Its chatbot is called Kimi and Alibaba is among its investors. K3 is its newest model. It has been answering API calls since 16 July. Today Moonshot released the model file itself.

A parameter is a number the model learned during training. K3 has 2.8 trillion of them, more than any model whose weights have been made public before.

It does not use all of them for any given piece of text. K3 is a mixture-of-experts model, which means it contains 896 specialist sub-networks. A routing layer picks 16 of them for each token it processes and leaves the other 880 idle. About 104 billion parameters are active at any moment. Moonshot’s model card confirms all of these numbers.

The routing explains why K3 runs at a reasonable speed. It does not reduce the storage requirement. Different tokens route to different experts, so a server has to hold the entire 2.8 trillion parameters in memory even though it only uses a fraction of them at a time.

K3 accepts 1,048,576 tokens of context, roughly 750,000 words in a single request. It reads images natively rather than handing them to a separate system.

Open weights and open source are different things

Moonshot has released the trained model. It has not released the training data or the code that produced it. You can download K3, run it, adapt it and charge people for access. You cannot rebuild it or audit how it was made.

The licence is not one of the standard open-source ones either. Most open models ship under well-known terms such as MIT or Apache 2.0, which lawyers already understand. Moonshot wrote its own. The weights are covered by a document it calls the Kimi K3 License, which permits commercial use but requires large companies reselling the model as a service to negotiate separately. Anyone planning to build a product on K3 should read those terms first.

What it costs

While downloading it is free, running it is not.

A 1.56TB set of weights needs that much fast memory before anything else is loaded. Moonshot recommends serving K3 on supernodes of 64 or more accelerators, and published deployment recipes point to configurations such as eight Nvidia B300 GPUs or 32 H100s with 80GB each. No laptop, workstation or single server can load it. Compressing it further does not help much, because Moonshot already trained the model at low precision rather than shrinking it afterwards.

So for nearly everyone, K3 is something you rent rather than own. Moonshot charges $3 per million input tokens and $15 per million output tokens. At Monday’s rate of about KES 129 to the dollar, that is roughly KES 388 and KES 1,940, before the 16% VAT Kenyan buyers pay on digital services. We covered Claude Opus 5 launching at $5 and $25 per million tokens on Friday.

K3 undercuts the American frontier models. It also costs about three times what Moonshot charged for its previous model. Chinese labs pricing at a tenth of everyone else is no longer the pattern.

The hallucination number, read correctly

One figure has travelled further than any other this week: K3 hallucinates 51% of the time. Several outlets translated that into roughly one in two answers being wrong. The benchmark does not say that.

The number comes from Artificial Analysis, an independent evaluation firm, on its AA-Omniscience test. Artificial Analysis defines the hallucination rate as the proportion of incorrect answers out of all non-correct responses, which it writes as incorrect divided by incorrect plus partial answers plus questions not attempted. The benchmark lets models decline to answer. The rate measures how often a model guesses when it should have declined.

On that test K3’s accuracy rose from 33% to 46% between generations while its hallucination rate rose from 39% to 51%. K3 knows more than its predecessor and bluffs more readily when it hits the edge of what it knows. Anyone using it for factual work should keep a person checking the output.

Low hallucination rates on this benchmark do not indicate a better model. The three lowest scores belong to MiniCPM5-1B at 1%, Command A+ at 14% and Grok 4.3 at 16%, which achieve them largely by declining to answer. Claude Fable 5, which leads the benchmark overall with an index score of 40 and 61% accuracy, does so by answering more questions correctly rather than by guessing less.

Why chip shares fell

Semiconductor stocks dropped when K3 launched. The Philadelphia Semiconductor Index fell into bear market territory on Friday 17 July, more than 20% below its late-June record. Global chip stocks had lost around $3.3 trillion in value since 22 June, a window that extends well before K3 appeared.

Investors were reacting to the possibility that a Chinese lab working under US export restrictions can approach the frontier without matching American GPU spending. That assumption supports a large part of the sector’s valuation.

Bloomberg’s analysts separated K3 from the DeepSeek shock of 2025. DeepSeek made models cheaper to train and run. K3 is far larger and leans on memory infrastructure, which supports demand for high-bandwidth memory rather than reducing it. Nvidia’s Jensen Huang argued on 22 July that more capable AI of any kind produces more usage, and more usage requires more computers. On 24 July, Nvidia, Microsoft, Meta, IBM and around 20 other companies signed a letter asking US lawmakers not to restrict open models.

The releases land on a financing structure we have written about before, in which Nvidia, OpenAI and Oracle fund each other’s growth. Pressure on model pricing at one end of that loop unsettles investors across all of it. Chip shares recovered part of the fall on 20 and 21 July.

The listings

Anthropic confidentially filed a draft registration statement with US regulators on 1 June. OpenAI announced its own confidential filing a week later, on 8 June, and said it had not settled on timing. Neither has listed.

Both companies will eventually have to explain to public investors why customers should pay for a closed model when a comparable open one can be downloaded. That question is now harder to answer than it was a fortnight ago.

Washington

Axios reported on 20 July that the Trump administration has revived plans to restrict Chinese AI models in the United States, through Entity List designations, federal procurement rules, security advisories and liability requirements for companies that host them. No executive order has been signed.

Enforcement is the difficulty. Once a 1.56TB file is downloaded and mirrored, no agency can withdraw it.

What this means in Kenya

Self-hosting K3 is not currently possible here. The requirement of 64 or more accelerators in one high-bandwidth cluster exceeds what local facilities offer. iXAfrica’s NBOX1 came online with 4.5MW of IT power, and Kenya’s largest planned AI facility, the Microsoft and G42 project, stalled on grid capacity.

The effect here is on price. A credible model at $3 and $15 per million tokens constrains what every other provider can charge, and Kenyan developers pay in dollars whichever model they choose. We argued in January 2025 that open models could widen access to AI. K3 is the strongest test of that argument so far, and the hardware requirement is the part of it that has not improved. For what is worth using today, our guide to the AI tools that matter in 2026 covers the field.

What happens next

Every performance claim about K3 until today came from Moonshot or from testing done through Moonshot’s own servers. Researchers can now run the model on their own hardware and check. Their results over the next few weeks will either match Moonshot’s published benchmarks or they will not. The other thing to watch is whether Washington signs anything. Both should be clear by the end of August.