Kimi K3 puts Moonshot’s largest open model push on the web and API

Moonshot AI has launched Kimi K3 across Kimi’s public web product and developer platform, positioning the model as its biggest step yet in the race to make frontier-class AI available outside closed U.S. lab systems.

The company’s Kimi site now presents “Kimi AI with K3” as an agentic workspace for coding and knowledge work. Its API platform says Kimi K3 is Kimi’s most capable flagship model to date, with a 1-million-token context window and use cases including software engineering, knowledge work, and deep reasoning. In a technical blog post, Moonshot describes K3 as a 2.8-trillion-parameter model with native vision capabilities, Kimi Delta Attention, Attention Residuals, and a sparse mixture-of-experts setup that activates 16 of 896 experts.

The practical news for users is that K3 is not only a paper launch. Moonshot says the model is available now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. The company lists API pricing at $0.30 per million tokens for cache-hit input, $3.00 per million tokens for cache-miss input, and $15.00 per million tokens for output. It also says full model weights will be released by July 27, 2026, with more architecture, training, and evaluation details due with a technical report.

That timing matters because K3 is being framed as an “open 3T-class” model, but the strongest current evidence is still mostly vendor-supplied. Moonshot says K3 is the first open model to reach 2.8 trillion parameters and that its full weights are coming later this month. Until those weights and the technical report are available, outside researchers cannot fully inspect training details, reproduce the company’s benchmark claims, or measure deployment costs across hardware configurations.

Moonshot’s own launch materials claim broad gains in agentic coding, multimodal work, and long-context knowledge tasks. The company says K3 can sustain long engineering sessions, navigate large repositories, coordinate developer workflows, and use screenshots or visuals in the loop to improve games, front-end work, and CAD-style tasks. It cites internal case studies in GPU kernel optimization, a “MiniTriton” compiler project, browser-based 3D game development, chip-design proof-of-concept work, and research workflows that combine literature review, code, equations, and interactive dashboards.

The API platform also makes the model legible as a product rather than only a research release. Its landing page lists K3 alongside K2.7 Code and K2.6, with K3 priced above those smaller or more specialized models. The page describes K3 as built for tasks “of any scale,” while the documentation index says the platform supports long context, multimodal understanding, and function calling. Those claims fit the current direction of AI product competition: vendors are trying to sell models less as chatbots and more as long-running agents for coding, research, office documents, and enterprise workflows.

Early outside signals are positive, but narrow. Arena posted that Kimi-K3 ranked No. 1 in its Frontend Code Arena and later No. 9 in its Text Arena, with gains from Kimi K2.6. The vLLM project also said it would support Kimi K3 on day zero and pointed to Moonshot work on efficient long-context serving. Those are useful indicators for developer adoption, but they do not substitute for independent safety evaluation, full-weight inspection, or controlled customer testing.

TechCrunch, citing a Financial Times report before the launch, said Kimi K3 was expected to be China’s largest open-weight AI model, with a parameter count between 2 trillion and 3 trillion, and part of a broader push by Moonshot to narrow the gap with closed-source systems from companies such as OpenAI and Anthropic. Moonshot’s published K3 materials now put the parameter count at 2.8 trillion and state that, in the company’s own view, K3 still trails the most powerful proprietary models while outperforming other models in its evaluation suite.

For developers, the immediate questions are practical: how well K3 performs outside vendor demos, whether the promised open weights arrive on time, how expensive the 1-million-token context is in real applications, and whether inference partners can make a model this large usable without a narrow set of high-end accelerator configurations. Moonshot’s technical blog says it recommends deploying K3 on “supernode” configurations with 64 or more accelerators, a reminder that open weights do not automatically mean cheap or simple self-hosting.

For readers tracking the AI race, K3 is another sign that the open-model lane is moving quickly toward frontier-scale systems. The model’s website launch gives everyday users and developers a way to try it now. The July 27 weight release, if it happens as promised, is the real test: that is when the claims move from a company-controlled product launch toward the harder world of independent replication, safety review, and deployment economics.

Sources

• Kimi K3 technical blog: “Open Frontier Intelligence”
• Kimi.com homepage for Kimi AI with K3
• Kimi API Platform
• Kimi API documentation index
• TechCrunch: Moonshot’s upcoming Kimi 3 is expected to close the gap with Anthropic’s Opus 4.8
• Arena post on Kimi-K3 in Frontend Code Arena
• vLLM project post on Kimi K3 support

---

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