Chinese startup launches world's largest open AI model 'Kimi K3', nears Fable and GPT-5.6 level performance

Chinese AI lab Moonshot has announced the largest open-weight AI model – Kimi K3. According to the company, Kimi K3 can match, if not outperform, the most advanced AI models from OpenAI and Anthropic. Here is all you need to know.

In Short

• Kimi K3 is the world's largest open AI model
• Moonshot says Kimi K3 gets close to Claude Fable and GPT-5.6
• Kimi K3 stood at first place in Frontend Code Arena benchmark

China is back at it again. Chinese AI startup Moonshot has unveiled Kimi K3, claiming it to be the largest open-weight AI model with 2.8 trillion parameters. As per the AI startup, Kimi K3 is a big improvement compared to previous versions, and now can match if not outperform the most advanced models from the US, including Claude Fable and GPT 5.6.

The Kimi K3 announcement comes at a time when the US is considering restricting advanced AI models made by US companies over fears of potential misuse. The US had temporarily blocked Claude Fable 5, and asked for a limited release of GPT-5.6.

What is Kimi K3?

Kimi K3 is an open-weight AI model. This means that anyone can freely download the model and run it locally on their device, making it cheaper than close-weight models like Claude Fable. To give you some context, Fable and GPT-5.6 can only be accessed via the cloud, and cost higher to run due to token pricing.

But open-weight models have existed for some time now, including Moonshot’s older models. What makes Kimi K3 different is how large it is. The model is trained on 2.8 trillion parameters – the more parameters, the better a model can perform. For context, as per reports, Anthropic’s Claude Opus 4.8 model was likely trained somewhere around 2 billion parameters.

How good is Kimi K3?

As part of the announcement, Moonshot shared benchmark results, comparing Kimi K3 to rival models. The startup said that Kimi K3 “performed competitively with Fable 5” and “substantially outperformed” Opus 4.8, GPT 5.6 Sol and GPT 5.5 in GPU kernel optimisation, a term used for methods that improve hardware use and reduce latency.

Third-party evaluations also placed the model close to top US systems. Arena AI ranked Kimi K3 first in its Frontend Code Arena. Arena AI confirmed that the AI model had managed to surpass Claude Fable 5, and jumped from 18th place for Kimi K2.6.

Vals AI placed Kimi K3 second overall behind Fable 5 and ahead of GPT-5.6 Sol, while Artificial Analysis said its performance was comparable to GPT-5.5 and Claude Opus 4.8, especially on complex multi-step tasks.

That is to say, that while Kimi K3 is not the most advanced AI model out there, it has closed the gap to frontier models from US companies. It is also the most-advanced open-weight model at the moment.

Moonshot says that Kimi K3 has a one million-token context window, supports native multimodal use, and is built for advanced reasoning, long-horizon coding and knowledge work, with the ability to handle long engineering sessions with limited human supervision.

Kimi K3 is now available on Kimi Work, Kimi Code, Kimi API, and Kimi.com. The company says that the open-weights will be made available by July 27.

Chinese AI models getting better?

The release also adds to a broader shift in which Chinese AI developers are moving faster and challenging the view in parts of the West that they remain behind US companies.

Kimi K3’s release comes as Chinese AI companies speed up model launches. Z.ai’s GLM-5.2 had already drawn attention for scoring close to leading US closed-source models. Hong Kong-listed MiniMax is said to be developing its own 2.7-trillion parameter model for release as early as the third quarter of 2026.

Before Kimi K3, Meituan’s LongCat-2.0 and DeepSeek’s V4-Pro, each with 1.6 trillion total parameters, had been among the largest Chinese models.

The release of Kimi K3 also signals that Chinese AI labs have been able to train advanced models despite restrictions from the US when it comes to sales of advanced GPUs to China. Earlier this year, Moonshot president Yutong Zhang said at the World Economic Forum, “We knew we didn’t have the luxury to simply scale up compute. That forced us to focus on fundamental research and efficiency.”

Do note that Anthropic has previously accused Chinese labs of carrying out “industrial-scale distillation attacks” against its AI models. Distillation refers to the use of outputs from larger models to train smaller ones. Though Chinese companies have rejected such claims.

The debate is not only about performance, but also about cost and adoption. Reports state that many companies in the West were turning to cheaper Chinese models to cut spending on US systems, whose access charges have risen this year. Cursor reportedly used Kimi in building Composer 2, while DoorDash delegated lower-level work to Kimi K2.6. Thinking Machines, the AI startup formed by Mira Murati, also used Kimi K2.5 to generate early post-training data for its Inkling model.

Backed by companies including Alibaba and Tencent, Moonshot has been expanding its capital base as it pushes to stay near the front of the sector. As per reports, the start-up was seeking about $2 billion in fresh funding at a valuation of roughly $30 billion ahead of a possible Hong Kong listing.

- Ends