Chinese AI startup Moonshot has unveiled the Kimi K3 model with 2.8 trillion parameters. According to the developers, it is the world's largest open-source AI model, and its performance is already approaching the advanced Fable model from American giant Anthropic.
Chinese AI startup Moonshot has unveiled the Kimi K3 model with 2.8 trillion parameters. According to the developers, it is the world's largest open-source AI model, and its performance is already approaching the advanced Fable model from American giant Anthropic.
Kimi K3 is the first open-source model to approach 3 trillion parameters. It is designed for complex logical reasoning, long-term coding and analytical work with large amounts of data, Reuters reports .
The model has a context window of 1 million tokens, which allows it to process and hold in memory significantly more data within a single query than previous generations of AI.
According to the company, in terms of GPU optimization, Kimi K3 “competed well with Fable 5 (including rollback) and significantly outperformed Anthropic’s Opus 4.8, as well as GPT 5.6 Sol and GPT 5.5.” This term refers to methods that make the most of AI hardware and minimize latency.
The model also demonstrated high results in independent third-party tests. In particular, the Arena.ai platform gave the Kimi K3 first place in a benchmark that assesses the capabilities of creating web interfaces.
Meanwhile, in the Vals AI ranking, the model took second place overall, behind only Fable 5, but ahead of GPT-5.6 Sol. Experts from the analytical company Artificial Analysis noted that the performance of the new product is quite comparable to OpenAI GPT-5.5 and Claude Opus 4.8 from Anthropic — especially in tests for performing complex multi-step tasks.
The huge size of Kimi K3 also means that it is unlikely that many users will be able to run it on their own servers, despite open access to the model's scales.
Ryan Fedasiuk, a research fellow at the American Enterprise Institute (AEI), noted in a LinkedIn post that running a model with 2.8 trillion parameters locally would require hundreds of thousands of dollars in computing hardware.
Moonshot also revealed that Kimi K3 has received two important architectural updates. They improve computational efficiency and allow the model to perform complex, long-running programming tasks with minimal human intervention.
