Jul 17, 2026 12:03:00

A Chinese-made AI model, 'Kimi K3,' comparable to GPT-5.6 Sol, has emerged, boasting a massive 2.8 trillion parameters in an open-source model.

Chinese company

Moonshot AI announced its AI model ' Kimi K3 ' on July 17, 2026 (Japan time). It has achieved scores exceeding GPT-5.6 Sol and Claude Fable 5 in some benchmark tests.

Kimi K3 Tech Blog: Open Frontier Intelligence
https://www.kimi.com/blog/kimi-k3

Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional… pic.twitter.com/eFHEbdxn3P

— Kimi.ai (@Kimi_Moonshot) July 16, 2026

Kimi K3 has 2.8 trillion parameters and a context window of 1 million tokens. Moonshot AI is promoting Kimi K3 as the 'world's first open 3T-class model,' and claims it will be effective in areas such as 'long-term coding,' 'knowledge work,' and 'inference.'

The benchmark test results for 'Kimi K3,' 'Claude Fable 5,' 'Claude Opus 4.8,' 'GPT-5.6 Sol,' 'GPT-5.5,' and 'GLM-5.2' are shown below. Kimi K3 recorded higher scores than the cutting-edge closed models Claude Fable 5 and GPT-5.6 Sol in multiple tests.

In coding performance tests, it demonstrated performance comparable to Claude Fable 5 and GPT-5.6 Sol.

The Kimi K3 is also characterized by its high cost-effectiveness. The graph below shows cost on the horizontal axis and the score from the web search performance test '

BrowseComp ' on the vertical axis, demonstrating that the Kimi K3 can achieve a high score at a low cost.

The third-party organization

Artificial Analysis has also published its performance analysis results for Kimi K3. In the Artificial Analysis Intelligence Index, which evaluates intelligence performance, Kimi K3 was rated as outperforming Claude Opus 4.8 and GPT-5.6 Terra.

It also outperforms Claude Opus 4.8 in the Artificial Analysis Coding Index, which evaluates coding performance.

In the Artificial Analysis Agentic Index, which evaluates agent performance, it also outperformed Claude Opus 4.8 and GPT-5.6 Terra.

The Kimi K3 announcement page also includes a demo video of a game created with Kimi K3, highlighting its ability to develop games that heavily utilize 3D models and physics simulations.

Kimi K3 is available for use in the chat AI 'Kimi.com' and the coding agent 'Kimi Code'. Furthermore, the entire model is scheduled to be released as an open model on July 27, 2026.

Furthermore, it is also available via API, with a fee of $3 (approximately 487 yen) per 1 million tokens for input, $0.30 (approximately 48.7 yen) for cached input, and $15 (approximately 2437 yen) for output.

Flagship Model Kimi K3 Pricing - Kimi API Platform
https://platform.kimi.ai/docs/pricing/chat-k3