Symbolic imageMoonshot releases Kimi K3, the largest open-weight AI model to date
Chinese start-up Moonshot AI has unveiled Kimi K3, a model with 2.8 trillion parameters whose weights are being published for free download from 27 July 2026. Company-run benchmarks place it ahead of the strongest closed models from Anthropic and OpenAI, and American developers and firms are already switching to it on cost grounds. Washington accuses Moonshot of building the model off Anthropic's technology; Beijing calls the charge groundless.
What happened
Moonshot AI, a Chinese start-up founded in 2023, presented its Kimi K3 model in the early hours of 16 July 2026, with the parameter weights to be made publicly available from 27 July, according to El País; Bloomberg likewise reports the model is being released for download. K3 was trained with 2.8 trillion parameters, which makes it the largest open-source model in the world, ahead of DeepSeek's largest model at 1.6 trillion. It processes one million tokens in a single task and uses a system that activates only the parameter subsets needed for a given query rather than the full model, an approach that reduces energy use per request. Performance figures published by the company place K3 ahead of Anthropic's Claude Fable 5 and OpenAI's GPT5.6 Sol; independently, K3 ranked first on the LLM Arena leaderboard for building online interfaces and scored at levels comparable to OpenAI and Anthropic on other software development tasks. Because the weights are public, any researcher, institution or company can download and adapt the model, as with DeepSeek's or Meta's Llama models. K3 follows Z.ai's GLM-5.2 in June, Alibaba's Qwen3.8 Max preview in July and DeepSeek's V4 previews in April.
The camps
The US administration, according to the Associated Press report carried by Daily Sabah, accused Moonshot on Wednesday of using "covert" but not necessarily illegal methods to build K3 off the back of Anthropic's Fable model. American politicians and AI companies including Anthropic allege illicit "distillation" of their models to extract the underlying technology; Beijing rejects that claim as "groundless". US Treasury Secretary Scott Bessent has warned that further sanctions could follow to protect American intellectual property, on top of existing American-led restrictions that block China from buying the most advanced AI chips. Moonshot itself frames K3 around efficiency, stressing that its full computing capacity is not used for every prompt. Inside the American market the reaction is not uniform: Mozilla's chief technology officer told AP he switched to K3 for much of his daily work within days of launch because it "just seems snappier" than the more expensive Claude Fable, and the exchange Coinbase says it is moving to Chinese models to cut costs. The Financial Times frames the open question as one of risk, asking what exposure users take on when they run Chinese open models such as K3.
The view from outside
European and Turkish outlets treat the release primarily as a shift in industrial position rather than a security incident. Handelsblatt places Moonshot alongside DeepSeek and Zhipu as Chinese labs pushing for the top of the global field. El País describes Silicon Valley as put on alert and highlights the commercial problem an open model of this quality creates, quoting a former Meta product director asking on X how Anthropic can justify Fable's pricing once the best open-weight model outperforms the best closed one. Daily Sabah, carrying AP, emphasises that Chinese models are cheaper and increasingly efficient and that, short of an outright ban, they are likely to keep appealing to independent developers in the US and elsewhere.
What's new
The dispute over open-weight models had until now run largely inside the American industry: Nvidia, Microsoft and IBM have endorsed distributing model parameters openly, while the New York Times reported on 25 July 2026 that Anthropic and OpenAI had broken with much of the rest of the sector on the question. Kimi K3 moves that argument outside the United States. For the first time the leading open-weight model is Chinese and, by its developer's own benchmarks, ahead of the leading closed models — and the American response has escalated from commercial argument to a government accusation about how the model was built.
What could happen next
Publication of the weights on 27 July makes the model permanently copyable, so adoption by developers and cost-sensitive companies outside China could continue regardless of what Washington decides; the reported switches at Mozilla and Coinbase would then be early examples rather than exceptions. Alternatively, the distillation accusations and the sanctions Bessent has signalled could harden into rules that restrict commercial use of Chinese models in the US, which would split the market between developers who can use K3 and those who cannot, without removing the downloaded weights already in circulation. A third path runs through pricing: if an open model matching or beating Claude Fable and GPT5.6 Sol is freely available, the closed-model vendors face pressure to justify subscription costs, and may respond by cutting prices, opening weights of their own, or competing on features that a downloadable model cannot supply.
China: Chinese labs present K3 as an efficient, openly downloadable model that competes on merit, and Beijing dismisses the distillation accusations as groundless.
United States: The administration and leading US AI firms say Moonshot built K3 using covert methods off Anthropic's technology and warn that further sanctions may follow.
Also involved