Enterprise

Moonshot releases Kimi K3 weights, dropping a 2.8-trillion-parameter open model into enterprise buyers' laps

The Beijing startup's July 27 weight release makes the largest open-weight model ever built free to download — and forces a rethink of what closed-model contracts are actually worth.

By Asher Glick · Enterprise software desk · 2026-07-27 · 3 min read

Photo: Unsplash / Taylor Vick — Rows of illuminated servers in a data center corridor

Moonshot AI pushed the full weights of Kimi K3 to HuggingFace on July 27, dropping a 2.8-trillion-parameter model, the largest open-weight system ever published according to Bloomberg and VentureBeat, into the public domain under a modified MIT license. The Alibaba-backed Beijing startup didn’t announce a product. It repriced a market.

The physical artifact is roughly 1.4 terabytes in MXFP4 quantization, with a one-million-token context window. Arena blind evaluations show developers picking K3 over leading U.S. models on front-end coding tasks, though Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol still sit above it in overall rankings. Independent testing from Artificial Analysis has K3’s hallucination rate climbing from the predecessor’s 39% to roughly 51%, a regression Moonshot hasn’t publicly addressed.

The commercial signal is louder than the benchmark noise. Moonshot’s ARR moved from $200 million in April to $300 million in June, and daily revenue has grown at least sixfold since K3 launched. The company is raising at a $50 billion valuation with a Hong Kong IPO possible this year. Giving the weights away is the growth strategy.

Washington isn’t amused. Michael Kratsios, director of the White House Office of Science and Technology Policy, has accused Moonshot of training K3 on banned Nvidia chips and running large-scale distillation against Anthropic’s models. The UK AISI and US CAISI issued a joint cyber risk assessment flagging the expanding cyber capabilities of open-weight systems and stating that zero-trust network proxies and strict execution sandboxes are mandatory requirements for enterprise API integrations.

And here’s the structural gap. Bloomberg Intelligence analysts note that none of the three major hyperscalers, AWS Bedrock, Azure Foundry, or Google Vertex AI, have integrated K3 or any comparable Chinese open-weight model. Enterprise buyers can download a frontier-adjacent model for free, but they can’t buy it through the procurement channels their compliance teams already trust.

That’s the real story of the release. The weights are legible to any engineering team with the storage to hold them; the geopolitical wrapper around them is illegible to every hyperscaler contract in force. Closed-model pricing has always assumed the alternative was worse. On July 27 that assumption stopped being automatic.

Sources