The Chinese AI lab Moonshot AI has officially released Kimi K3, its newest AI model which it said is as good or better than the leading models from OpenAI and Anthropic across a range of benchmarks.
In its announcement Moonshot AI scored 81.2 per cent on the coding benchmark FrontierSWE, ahead of all other AI models except Claude Fable 5.
The largest open-source model ever released, at 2.8 trillion parameters, Kimi K3 will be accessible via Moonshot AI’s API, in a private cloud, or on premises.
Independent AI analysis firm Artificial Analysis ranks Kimi K3 third overall in its intelligence index, behind only GPT-5.6 Sol and Fable 5. The index is a composite of nine AI benchmarks across mathematics, science, coding, and reasoning.
The crowdsourced platform Arena AI, which ranks AI model performance based on blind A/B testing, now ranks Kimi K3 as the best model for front-end coding, ahead of both GPT-5.6 Sol and Fable 5. Its front-end leaderboard considers performance in tasks such as data analysis, marketing, simulations, and consumer-facing product.
Artificial Analysis also noted that Kimi K3 is more expensive than other open source AI models, at $0.94 per task on average compared to $0.47 per task using Z.ai’s GLM-5.2 or $0.04 per task using DeepSeek V4 Pro.
Though Nvidia and the US government believe that advanced GPUs are being smuggled into China, Chinese AI firms do not officially have access to the latest chips. As a result, Chinese AI labs have had to make their models as efficient as possible to train and run on less powerful hardware.
Moonshot said it had rewritten the architecture that underpins Kimi K3, allowing it to train the massive model more efficiently, as well as a new Mixture of Experts (MoE) approach that breaks the model up into 896 separate specialised ‘agents’, only 16 of which are active at any given time.
Western AI labs have also accused Chinese developers of “distillation”, a process by which open-source AI models are trained using the outputs of competitor platforms such as ChatGPT or Claude.
In February, Anthropic accused Moonshot AI of using “hundreds of fraudulent accounts” for 3.4 million exchanges with Claude. The firm said it believed Moonshot staff had attempted to distil Claude’s agentic reasoning, coding capabilities, and computer vision functions.
