Moonshot AI drops open-source Kimi K2 with agentic superpowers

Kimi K2 is an open-source mixture-of-experts model from Moonshot AI, packing 1 trillion total parameters with 32 billion activated during inference. Instead of behaving like a standard conversational chatbot, it is tuned specifically for autonomous agents and programming tasks, built to write code, call tools, and finish multi-step workflows without constant human prompts.

Moonshot released two versions of the model: a foundation model for developers and researchers, and an instruction-tuned model named Kimi K2-Instruct optimized for interactive chat and autonomous agent pipelines.

Benchmarks and agent capabilities

On the real-world LiveCodeBench evaluation, Kimi K2 scored 53.7% accuracy, comfortably beating DeepSeek-V3 at 46.9% and GPT-4.1 at 44.7%.

For autonomous software engineering, it posted a 65.8% solve rate on SWE-bench Verified, matching several prominent proprietary models.

It also showed strong mathematical reasoning on MATH-500, hitting 97.4% compared to 92.4% from GPT-4.1.

Stability via the MuonClip optimizer

Training trillion-parameter models usually comes with severe training instabilities that derail expensive cluster runs.

Moonshot tackled this using an in-house optimizer dubbed MuonClip. By departing from standard AdamW variants and rescaling weight matrices in query and key projections, the team reported completing the training run with zero instability crashes.

Alongside the open-source release, Moonshot made the model accessible via an aggressive API tier. Cache-hit input tokens are priced at $0.15 per million, while output tokens cost $2.50 per million.

Why it matters

If you are building autonomous coding agents or complex tool-use systems, Kimi K2 gives you frontier-level code execution without being chained to proprietary endpoints.

We think this launch accelerates the commoditization of agentic capabilities. By delivering competitive coding benchmarks under an open architecture and pairing it with rock-bottom API pricing, Moonshot is squeezing the high margins that closed-source labs have relied on.