Chinese artificial intelligence (AI) startup Moonshot AI unveiled its next-generation AI model, "Kimi K3," which has 2.8 trillion parameters. Independent evaluations said it posted performance close to the top models from U.S. Anthropic and OpenAI, fueling assessments that Chinese AI corporations are rapidly narrowing the technology gap.
According to the industry on the 19th, Moonshot AI recently unveiled the AI model Kimi K3. Released a day before the World Artificial Intelligence Conference (WAIC 2026), which opened in Shanghai on the 17th, Kimi K3 features native multimodal capabilities that accept images and text together. It is characterized by a context window that can process up to 1 million tokens at once. It is designed for extended coding and knowledge work and complex reasoning. Moonshot AI described Kimi K3 as the world's first open-weight AI model at the 3-trillion-parameter level.
Parameters are internal variables that AI learns during training and are an indicator of model size. In general, more parameters can allow the model to learn more complex information, but they do not necessarily correlate with actual performance.
An open-weight model allows developers to download the trained AI's core weights and run it on their own servers or modify it for specific uses. However, Moonshot AI is currently offering Kimi K3 through its own application programming interface (API) and has not yet released the full model weights. The company plans to release the weights later.
Kimi K3 scored 57 on the intelligence index assessed by AI performance analysis firm Artificial Analysis, ranking third overall. It trailed Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, but is on par with Claude Opus 4.8 and GPT-5.5.
It also received high marks for AI agent tasks that perform coding and multiple steps autonomously. In Arena AI's web interface development evaluation, it ranked first, and in Vals AI's comprehensive evaluation, it placed second after Fable 5, ahead of GPT-5.6 Sol. However, benchmark scores can differ from response quality and stability in real service environments.
Moonshot AI said Kimi K3 is suitable for long-running tasks with reduced human intervention, such as analyzing large software repositories and operating terminal tools. It also supports creating and editing digital content, including webpages and games, based on image inputs.
Pricing is higher than existing Chinese open models. Kimi K3's API fees are $3 for input and $15 for output per 1 million tokens. Artificial Analysis estimated the average expense per task at $0.94. It is similar to GPT-5.6 Sol and cheaper than Claude Opus 4.8, but more expensive than other Chinese open models such as DeepSeek V4 Pro and GLM-5.2. This is seen as a move away from the ultra-low-price strategy that Chinese AI corporations have typically emphasized, aiming to secure profitability with high-performance models for enterprise coding and knowledge work.
Given the model's size, it will not be easy for users to operate it themselves. Experts analyzed that running a 2.8-trillion-parameter model on private servers could require computing equipment worth hundreds of thousands of dollars. Even if the weights are released, many users are likely to use services from Moonshot AI or cloud providers.
Industry observers say China's open AI ecosystem is rapidly narrowing the gap with the top U.S. closed models. That is because, following last year's DeepSeek, Kimi K3 is also demonstrating high performance.