Kimi says K3 is a 2.8T-parameter AI model built for long-context agent work

Dek: Moonshot AI’s Kimi site says K3 is live on the web, in Kimi Code and through its API, with a 1-million-token context window, native vision and a full-weights release planned for July 27.

Moonshot AI’s Kimi has launched K3, a new large language model that the company is positioning as an “open frontier” system for coding, long-context knowledge work and multimodal agent tasks.

The company’s Kimi website now promotes “Kimi AI with K3,” and Kimi’s technical blog says K3 is available today on the Kimi app, Kimi Work, Kimi Code and the Kimi API. The API documentation also lists K3 as live and describes it as Kimi’s most capable model to date.

Kimi’s central claim is scale. The company says K3 has 2.8 trillion parameters, uses Kimi Delta Attention and Attention Residuals, supports native visual understanding, and can handle a 1-million-token context window. Kimi says the model is designed for “frontier intelligence scenarios” including software engineering, knowledge work and deep reasoning.

The release is aimed squarely at the agentic coding and work-assistant market. Kimi’s product pages pitch K3 for building playable multiplayer and 3D games, creating consulting-style slides, handling large codebases and running parallel tasks. Kimi Code’s documentation says K3 is available in the coding product, with 1M context available to higher-tier members.

The important caveat: several of the biggest claims are still company-reported. Kimi’s benchmark tables, case studies and “frontier-level” language come from its own release materials, not from a broadly replicated independent test suite. The company itself says K3 still trails the most powerful proprietary models overall, while claiming it outperforms other tested models across parts of its evaluation suite.

Availability is also layered. Kimi says the model is usable now through its consumer and developer surfaces, including kimi.com, Kimi Work, Kimi Code and the Kimi API. But the full model weights are not yet published. Kimi says it plans to release them by July 27, 2026, alongside more technical detail on architecture, training and evaluations.

For developers, the API quickstart lists the model name as kimi-k3 and says K3 supports vision input, custom tools and tool choice, dynamic tool loading, automatic context caching and official tools. The pricing Kimi gives in its blog is $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens and $15.00 per million output tokens.

The launch also arrives with a documentation wrinkle worth spelling out. Kimi Code’s release notes describe K3 as “released and open-sourced,” while Kimi’s technical blog says the full model weights will be released by July 27. That is not just a wording issue for developers. Public copy should avoid telling readers the weights are already available unless the actual release page, license and downloadable artifacts are visible. The safer phrasing is that K3 is live as a product and API today, with open weights promised on a dated timeline.

Kimi’s release materials lean heavily on agent examples: autonomous kernel optimization, a compact Triton-like compiler, a browser-based 3D game, a chip-design proof of concept, financial-research dashboards and scientific-computing workflows. Those examples are useful signals about the market Kimi wants to win — not proof that typical users will get the same results. The story should keep the examples in the “company says” lane unless Shadowfetch or an independent lab reproduces them.

There is some outside context already. Artificial Analysis has a live Kimi K3 page and describes the model as among the leading models in intelligence, while also flagging that it is slower than average and verbose. That supports a measured framing: K3 is being watched by independent evaluators, but the outside evidence available now does not settle the strongest claims in Kimi’s launch post.

That combination — a very large sparse model, a million-token context window, native vision and promised open weights — is why the launch matters. If the open release arrives on schedule and independent evaluators can reproduce meaningful parts of Kimi’s performance claims, K3 could become a serious option for developers and companies that want frontier-class agent workflows without depending entirely on closed Western model providers.

For now, the clean read is narrower: Kimi has launched K3 on its website and API, is marketing it around long-horizon coding and knowledge work, and has set a near-term date for open weights. The performance story needs independent testing before it should be treated as settled.

Sources

• Kimi K3 technical blog: “Open Frontier Intelligence”
• Kimi AI homepage with K3 positioning
• Kimi K3 API quickstart
• Kimi Code “What’s New” documentation
• Artificial Analysis model page for Kimi K3

---

Shadowfetch is a technology publication. Explore Shadowfetch Linux — our own Linux build — and the Shadowfetch apps on the App Store.