Moonshot AI launches Kimi K3 to rival US frontier models

Moonshot AI released Kimi K3 on 16 July 2026, a 2.8 trillion-parameter model the company describes as the world's first open AI system in the three-trillion-parameter class, putting it on a par with the leading models from Anthropic and OpenAI.

The Beijing-based start-up made the model available immediately through its Kimi app and API, with full open weights scheduled for public release on 27 July. Once published, any developer can download, run, and modify the system without restriction, a capability that closed models from Anthropic's Claude and OpenAI's ChatGPT do not offer.

Kimi K3 uses a sparse Mixture-of-Experts architecture, meaning only a fraction of its 2.8 trillion parameters activate on any given request. That design cuts inference costs compared to a dense model of equivalent size, allowing Moonshot to price API access at $3 per million input tokens and $15 per million output tokens. Cached input tokens drop to $0.30 per million.

The model supports a 1-million-token context window and accepts text, image, and video input. Moonshot says it is built for long-horizon tasks including software engineering, coding, and knowledge work, and is designed to operate with minimal human supervision across extended autonomous runs.

Third-party evaluations from Artificial Analysis and Arena.ai placed K3 at an overall Elo score of 1,547, trailing only two proprietary American models in the rankings. On Arena.ai's frontend code benchmark, K3 ranked first with a score of 1,679 points, placed above Anthropic's Fable 5 in blind developer testing. Moonshot's own benchmarks show K3 outperforming Claude Opus 4.8 and GPT-5.5 on most tasks, while falling below Fable 5 and GPT-5.6 Sol at the very top.

The pricing structure deserves particular attention. At $3 per million input tokens, K3 costs roughly half what Anthropic charges for comparable Opus-tier access. For development teams running high-volume workloads, that gap compounds quickly. Several developers who tested K3 on real coding tasks reported in public forums that they could not distinguish its output from Anthropic's Fable 5.

The model comes in two variants. K3 Max targets chat and agent workloads. K3 Swarm Max is designed for large-scale parallel processing, where multiple instances run simultaneously across independent tasks. Both variants were available from launch on 16 July through Kimi Code, the Kimi app, and the API.

Independent benchmark confirmation of Moonshot's own figures is still pending. The open-weight release on 27 July will give the broader research community its first opportunity to run the full model, verify the parameter count, and stress-test the claimed performance on a wider range of tasks.

The announcement had an immediate effect on Chinese AI stocks. Shares in domestic competitors Zhipu and MiniMax fell roughly 27% and 16% respectively in Hong Kong trading on the day of the release, as investors reassessed the competitive positions of the two firms in light of K3's reported capabilities.

Moonshot is backed by Alibaba, Tencent, and Meituan, among others. The company closed a $2 billion funding round in May 2026 at a $20 billion valuation and has since held early talks with new investors at a reported target of $30 billion, according to people familiar with the discussions. That trajectory marks a sevenfold increase from its December 2024 valuation of $4.3 billion. A previous funding round valued the company at $4.8 billion in January 2026.

The K3 release fits a broader pattern across China's AI sector. Labs including DeepSeek, Zhipu, and MiniMax have each produced models that challenge American equivalents on benchmarks, doing so under US export controls that restrict access to the most advanced Nvidia chips. Moonshot's own disclosures note the use of export-grade Nvidia silicon alongside an unnamed alternative GPU supplier, a configuration that has become standard across the sector.

The open-weight release on 27 July will be the clearest test of K3's practical reach. Developers who download the weights can host the model on their own infrastructure, bypassing API costs entirely, and modify it for specialist applications. The last open model to shift developer workflows at scale was DeepSeek R1 in early 2025. Moonshot is betting K3 will be the next.