Kimi K3 is here: 2.8 trillion parameter open-source AI model rivals GPT and Claude

Moonshot AI just changed the open-source AI landscape. On July 16, the Beijing-based startup released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that is now the largest open-weight model in existence. It performs competitively with GPT-5.6 and Claude Opus 4.8 on key benchmarks, and the full weights will be publicly available by July 27.

This is a significant moment. For the past 18 months, DeepSeek dominated the open-source AI conversation with efficient, cost-effective models. Kimi K3 is Moonshot AI's response: not cheaper, but bigger. The bet is that raw scale still matters, and that the developer community will rally around a model that matches frontier proprietary systems without the API lock-in.

What makes Kimi K3 different

K3 is built on a sparse MoE architecture with 896 experts, of which 16 are activated per token. This selective activation keeps inference costs manageable despite the massive parameter count. The model also introduces Kimi Delta Attention and Attention Residuals, two architectural innovations that improve training efficiency and long-context performance.

The context window is 1 million tokens, putting it in the same league as Gemini and Claude for long-document tasks. It also handles images and video natively, no separate vision encoder needed.

Benchmark performance

In Moonshot's published evaluations, K3 leads on BrowseComp, SWE Marathon, and OmniDocBench. It trails Fable 5 on FrontierSWE and HLE-Full, but beats both GPT-5.6 Sol and Claude Opus 4.8 across most knowledge work and coding benchmarks.

Pricing and availability

K3 is live now on Kimi.com, Kimi Work (desktop app v3.1.0+), Kimi Code (terminal), and the Kimi API. The API is OpenAI SDK-compatible, so switching from GPT or Claude requires changing only the base URL. Pricing is $3 per million input tokens, $15 per million output tokens, with cached input at $0.30 per million. A promotional rebate running through August 12 gives up to 30% back in vouchers for API credits of $1,000 or more.

The model weights will be released on July 27. Until then, developers can test the model through the API or the chat interface at kimi.com. No credit card is required to start chatting.

What this means for the AI landscape

K3 represents a strategic shift for Moonshot AI. After DeepSeek's R1 disrupted the Chinese AI market in early 2025, Moonshot's market position eroded significantly. K3 is a calculated gamble: release the biggest open model in the world, reclaim the developer community's attention, and make it harder for enterprises to justify closed-source API spend.

For developers and teams evaluating AI models, K3 adds a powerful new option to the open-source stack. Combined with existing tools like Claude for writing and ChatGPT for general assistance, K3 specializes in long-horizon coding and deep knowledge work. We will publish a full hands-on review once the weights are available for local testing.