The model uses a mixture-of-experts design with a trillion total parameters but activates just 32 billion at any given time, making it surprisingly efficient. Moonshot is offering a base foundation model for researchers and developers, and an instruction-tuned variant aimed at chatbots and autonomous agent tasks.

Moonshot said in an announcement which clearly was written by an AI: ““Kimi K2 does not just answer; it acts. With Kimi K2, advanced agentic intelligence is more open and accessible than ever. We can’t wait to see what you build.”

The big selling point is “agentic” behaviour which is the ability to autonomously use tools, write and execute code, and pull off complex multi-step tasks without someone hovering over it. On SWE-bench Verified, a brutal software engineering benchmark, Kimi K2 hit 65.8 per cent accuracy, outclassing almost every other open model and even giving proprietary ones a run for their money.

It did better on LiveCodeBench, a coding benchmark designed to mimic real-world scenarios, scoring 53.7 per cent compared to DeepSeek-V3’s 46.9 and GPT-4.1’s lacklustre 44.7. It nailed 97.4 per cent on MATH-500, an advanced mathematical reasoning test, while GPT-4.1 lagged behind at 92.4 per cent. Moonshot might have stumbled onto something fundamental about mathematical reasoning that the usual suspects haven’t cracked.

What makes this sting for OpenAI is the cost. Kimi K2 was trained and optimised for a fraction of the eye-watering sums it is blowing on GPUs and compute for marginal gains. Moonshot’s leaner approach suggests the industry’s supposed “scale or die” narrative isn’t as bulletproof as it looked.