While American AI labs were busy price-warring each other into the ground, Moonshot AI quietly dropped a 2.7-trillion-parameter model, slapped a free tag on it, and walked off like it wasn't a big deal. It was a big deal.

The Biggest Open Weights on the Block

Kimi K3, released by Moonshot AI with open weights published on Hugging Face on July 26, is now the largest openly available large language model, edging out DeepSeek V4's 1.6 trillion parameters by a wide margin. It's built as a Mixture-of-Experts model, meaning it only activates a slice of those parameters per request, and it ships with a 1-million-token context window plus native image (and reportedly video) understanding.

You can try it free right now at kimi.com or in the Kimi app, no developer account required, and the open weights mean anyone with enough GPU horsepower can run it locally. For everyone else, the API runs $15 per million output tokens — a fraction of what the big Western labs charge for their flagship models.

Beating the Curve, Not Just the Benchmark

Moonshot claims K3 goes toe-to-toe with Anthropic's Fable 5 and substantially outperforms Anthropic's Opus 4.8 and OpenAI's GPT-5.6 lineup on released benchmarks, particularly in coding tasks, with developers in blind evaluations reportedly preferring its front-end code over both companies' offerings. Whether or not it's truly best-in-class across the board, it's arriving months ahead of when analysts expected China to close this particular gap.

The real story isn't just one good model — it's that "open, free, and genuinely competitive" used to be a pick-two situation for large language models, and Kimi K3 is quietly arguing you can have all three.

Somewhere, an AI lab's pricing team is having a very long meeting right now.

Source: Fortune