Beijing-based Moonshot AI's open-weight model scored 1,679 points, leapfrogging Anthropic and OpenAI in a ranking that matters for the AI-crypto crossover economy
An open-weight AI model built in Beijing just outperformed the flagship products of the two most valuable AI companies on the planet. Kimi K3, developed by Moonshot AI, debuted at the top of Arena.ai’s Frontend Code Arena leaderboard on July 16 with a score of 1,679, beating Anthropic’s Claude Fable 5 at 1,631 and OpenAI’s GPT-5.6 Sol at 1,618.
What Kimi K3 actually did #
The Frontend Code Arena isn’t some synthetic benchmark cooked up in a lab. It uses Elo-style voting on real-world coding tasks, meaning actual developers pit models against each other and vote on which output is better.
Kimi K3 didn’t just win overall. It topped six out of seven sub-domains assessed, including Brand & Marketing and Data & Analytics. The only category where it didn’t finish first still saw it competitive enough to hold the aggregate lead by a comfortable margin.
Moonshot AI’s previous model, Kimi K2.6, sat at 18th place on the same leaderboard. K3 vaulted 17 spots to claim the crown.
Under the hood, Kimi K3 runs on roughly 2.8 trillion parameters using a Mixture-of-Experts architecture. It also ships with a 1M-token context window, meaning it can process roughly 750,000 words of code or documentation in a single prompt.
Kimi Delta Attention and Attention Residuals were specifically designed to handle long-horizon coding tasks. Moonshot AI plans to release the full model weights by July 27, making it freely available for anyone to run, modify, or build on.
The pricing undercut #
Kimi K3 is available via API at $3 per million input tokens and $15 per million output tokens.
Why crypto should pay attention #
Moonshot AI, founded in Beijing with its initial Kimi chatbot launching in October 2023, is a pure AI play. There’s no token, no on-chain integration, no partnership with a Layer 1.
An open-weight model beating proprietary ones is especially relevant for decentralized AI projects. Protocols that rely on open models for inference, training, or fine-tuning now have access to a state-of-the-art coding model without licensing restrictions.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our