Kimi K3's open weights went live on Monday, making the largest open-weight model ever published free to download. Loading it takes roughly 1.4 terabytes. AWS Bedrock, Azure Foundry and Google Vertex AI carry none of it, and the White House says the model was trained on banned Nvidia chips.

On Monday, July 27, a Beijing startup published 2.8 trillion parameters to Hugging Face and walked away.

No paywall. No waitlist. No API key. The largest open-weight model ever released, free to anyone with the bandwidth to pull it and the hardware to hold it.

The hardware is the catch, and it is a large one. In the MXFP4 four-bit format Moonshot shipped, the weights alone occupy about 1.4 terabytes. At 16-bit precision the same model would need roughly 5.6 terabytes. Either number puts a full self-hosted deployment out of reach for a single machine and into the territory of a multi-node GPU cluster.

Ten days ago we reported that Kimi K3 had beaten Claude Fable 5 on front-end coding in blind developer evaluations and that the weights were not out yet. They are out now. The interesting part is who has not picked them up.

Amazon, Microsoft and Google Have Integrated None of It

Bloomberg Intelligence analysts Mandeep Singh and William Tong noted that none of the three major hyperscale platforms carry Kimi K3 or any comparable Chinese open-weight model. Not AWS Bedrock. Not Azure Foundry. Not Google Vertex AI.

For a model this capable, that absence is a market. Cloud providers that specialize in serving third-party open models inherit demand the hyperscalers are declining, and they inherit it on day one rather than after a procurement cycle.

The practical effect for an engineering team is a fork in the road. Serving K3 means either standing up your own cluster, or routing to a specialist host, or calling Moonshot's API directly. It does not mean adding a line to an existing Bedrock config.

The Case That This Is Logistics, Not Politics

The obvious reading of the hyperscaler gap is geopolitical. It is not the only reading available.

Hosting a 2.8-trillion-parameter Mixture-of-Experts model is a serving problem before it is a policy problem. A managed platform has to reserve multi-node capacity, build routing for 896 experts, negotiate indemnification, and publish a support commitment on a checkpoint whose technical report has not shipped and whose license terms are still being read. That work takes weeks regardless of where the weights came from, and these weights went public this week.

What keeps the political reading alive is silence. None of the three companies has said publicly why K3 is absent from its catalog. Until one of them explains, an engineering team cannot tell whether it is waiting on a procurement queue or on a policy decision, and those two things resolve on very different timelines.

The Architecture Is Why 2.8 Trillion Parameters Is Affordable at All

Kimi K3 is a Mixture-of-Experts model: instead of running every parameter on every token, the network routes each token to a small subset of specialized subnetworks. K3 holds 896 experts and activates 16 of them per token, which works out to roughly 50 billion parameters of live compute per step.

That ratio is the entire economic argument. A 2.8-trillion-parameter model that only lights up about 50 billion parameters at a time costs far less to serve than the headline number suggests.

Specification	Kimi K3
Total parameters	2.8 trillion
Experts	896, with 16 active per token
Active parameters per token	roughly 50 billion
Context window	1,000,000 tokens
Weight format	MXFP4, via quantization-aware training
Weight storage	approximately 1.4 TB
Modalities	text and native vision

The quantization detail matters more than it looks. Moonshot applied quantization-aware training rather than compressing a finished model afterward, so K3 learned to work around four-bit precision during training instead of absorbing the damage at the end. MXFP4 also runs natively on NVIDIA Blackwell and AMD MI400 hardware, which means the format is a deployment decision as much as a compression one.

A technical report covering the design, training and benchmark methodology is expected to follow the weight release.

The Debut Erased Billions From Its Competitors

The reaction when K3 launched was not measured in downloads.

Z.ai stock lost as much as 30% of its value in Hong Kong trading. MiniMax Group fell as much as 16%. Alibaba dropped 4%. Investors read one Chinese lab's frontier release as a direct subtraction from every other Chinese lab's future.

In Arena blind evaluations, developers picked Kimi K3 ahead of leading U.S. models on front-end coding tasks. On overall rankings it placed below Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, while beating both companies' previous-generation models across coding and general agent evaluations.

That is the specific shape of the current gap: the newest closed U.S. models still lead, and last year's closed U.S. models no longer do. We tracked the same crossover earlier this month, when Chinese open models overtook America's and Claude fell to seventh.

The Money Followed Immediately

Moonshot's daily revenue has grown by a factor of at least six since K3 launched.

Annual recurring revenue reached 300 million dollars in June, up from 200 million in April. The company is now raising at a 50 billion dollar valuation ahead of a possible Hong Kong listing that could come as soon as this year.

Giving away the weights and selling the API is not a contradiction. Most organizations that download 1.4 terabytes of weights will conclude that operating an eight-node cluster is someone else's job, and Moonshot is that someone.

Washington Says the Model Was Built on Banned Silicon

The release arrives under an accusation that Moonshot has not answered.

Michael Kratsios, Director of the White House Office of Science and Technology Policy, accused Moonshot last week of training K3 on banned Nvidia chips and of running large-scale distillation against U.S. models, including Anthropic's Fable. Distillation is the practice of using a stronger model's outputs as training data for your own, and it is the same accusation Anthropic brought to the Senate when it documented 25,000 fraudulent accounts probing Claude.

Moonshot has not responded to requests for comment on the allegations.

Neither claim has been publicly substantiated with evidence, and the open weights complicate rather than settle the question. Researchers can now inspect the model directly, probe how its 896 experts specialize, and test whether its behavior carries the fingerprints of any particular teacher model. That work has not been done yet. It is now possible, which was not true last week.

What Changes for Engineering Teams This Week

Nothing about your stack breaks because a Chinese lab published a checkpoint. A few things become worth checking.

The self-hosting math is now real, not theoretical. A 1.4 TB weight file with a 1M-token context is a capacity planning exercise, not a download. Budget for MoE-aware serving, which in practice means vLLM, SGLang or TensorRT-LLM with routing support for 896 experts.
Data-sovereignty arguments now have a frontier-class option. Teams that could not send code to a U.S. API endpoint previously had to accept a capability gap. That gap narrowed this week.
Procurement gets harder, not easier. With no hyperscaler carrying the model, using K3 inside a large enterprise means a new vendor relationship or new infrastructure, and both take longer than a benchmark score suggests.
The license and the technical report are still pending detail. Read both before committing a roadmap to this model.

The Bottom Line

A model that beats last year's best closed American systems is now sitting on a public download page, licensed for anyone to modify, and not one of the three American hyperscalers carries it.

Those two facts are in tension, and the tension is the story. The technical barrier to using frontier-class Chinese models fell to zero this week. The institutional barrier did not move at all. One of those is easier to fix than the other, and it is not the one involving 1.4 terabytes.

Kratsios says the model was trained on chips it was never supposed to have. Moonshot has said nothing. The weights, meanwhile, are already on thousands of hard drives, and no export control written so far reaches a file that has finished downloading.