Why everyone’s panicking over Chinese AI

Moonshot AI released the weights for Kimi K3 on July 27, making a 2.8 trillion parameter model, the largest openly available anywhere, free to download.

The specs explain the noise. K3 is a mixture of experts model that fires only 16 of its 896 experts per token, supports a one million token context window, and handles text, images and video natively. It scores 57 on the Artificial Analysis Intelligence Index, putting it in the top handful of models overall. No open model has ever ranked this high on an independent evaluation.

What open weights mean

Open weights mean the trained parameters are public. This is not open source: you get the finished model, not the training code or data. That is enough to self host and escape vendor lock in. It is also enough that capability, once released, cannot be recalled.

The accusation

White House OSTP director Michael Kratsios alleged that Moonshot trained K3 using export controlled Nvidia chips and through large scale distillation against American models, naming Anthropic's Fable. Moonshot has not responded. Independent analysts say K3's original architecture means distillation alone cannot account for its performance.

The asterisk on the benchmarks

On the evaluation where K3's accuracy climbed from 33% to 46%, its hallucination rate rose from 39% to 51%. The index score still improved, because the formula rewards accuracy gains more heavily than it penalises hallucinations. The model got more answers right and more answers confidently wrong at the same time.

Why the cost story is complicated

K3 undercuts Western frontier models on token price. Artificial Analysis also calls it notably slow and very verbose: 32 output tokens per second against a median of 72, and 160 seconds to first token. Cost per task is $0.94, close to GPT-5.6 Sol's $1.04 and higher than other open weight models. Self hosting needs 1.4 TB resident, meaning eight to sixteen nodes of H100 class GPUs. None of AWS Bedrock, Azure Foundry or Google Vertex AI offer it.

The policy fight

Axios reported on July 20 that the Trump administration is reviving a push to restrict Chinese AI models, and Commerce has weighed adding Chinese labs to its Entity List. Hugging Face, Meta, Microsoft, Mistral and Nvidia signed an open letter opposing broad restrictions. Beijing is running the argument in reverse: the Financial Times reported on July 21 that China's commerce ministry is consulting Alibaba, ByteDance and Zhipu on export controls of its own.

What it means for India

Cheap near frontier capability lowers the biggest barrier for Indian startups building on large models. But few Indian firms have the clusters to self host 1.4 TB, so most adoption will run through APIs and GPU cloud providers, not real in house deployment. The data sovereignty case only holds if you can host the thing.

The weights are out. Benchmarks stop mattering and production results start.