Moonshot releases open Kimi K2 Thinking: 300 tool calls in a row

Moonshot AI has released Kimi K2 Thinking, its strongest open reasoning model. It is built as an agent: it thinks step by step and calls tools between steps, making up to 200–300 calls in a row without human intervention. On Humanity's Last Exam with search, Python and a browser, the model scored 44.9%. It reached 60.2% on BrowseComp, where the human level is 29.2%, and 71.3% on SWE-Bench Verified. The model is available in the chat at kimi.ai and through the API.

How the agentic loop works

Moonshot calls this test-time scaling: it grows not only the number of reasoning tokens but also the number of tool-use steps. A typical loop looks like this: think, search, open a browser, think again, write code. Along the way, the model proposes and refines hypotheses, checks sources and assembles an answer. In one of the company's examples, it solved a PhD-level math problem in 23 interleaved steps of reasoning and tool calls.

The release's own footnotes contain a practical caveat. The chat at kimi.ai uses a reduced set of tools and fewer turns, so benchmark figures may not reproduce there. The full agentic mode in the product is promised for later.

Speed: INT4 without quality loss

The key engineering detail is quantization. Models with long reasoning generate huge sequences, and low-bit compression usually hurts quality noticeably. Moonshot applied Quantization-Aware Training during post-training: the weights of the MoE components are stored in INT4 (weight-only). This enables native INT4 inference and roughly twice as fast generation. All reported results were obtained in INT4, meaning there is no discount relative to "full" precision.

Key figures from the release:

• HLE with tools: 44.9% (51.3% without blocking access to Hugging Face).
• BrowseComp: 60.2%, human baseline 29.2%.
• SWE-Bench Verified: 71.3%; SWE-Multilingual: 61.1%; Terminal-Bench: 47.1%.
• Context: 256k tokens, temperature 1.0 in most tests.
• HLE: up to 120 steps and 48k reasoning tokens per step; agentic search: up to 300 steps and 24k tokens per step.
• Heavy Mode: eight parallel trajectories with subsequent aggregation of answers; GPT-5 Pro is used in the comparison.

What to check in the methodology

Moonshot discloses the test conditions in detail, which lets readers assess the numbers critically. For HLE, access to Hugging Face was blocked because of the risk of answer leakage; with access, the result would rise to 51.3%.

The grader was o3-mini with the official HLE prompts. For GPT-5, a high reasoning effort level was used. Some of the competitors' results come from official publications (on HLE with tools, GPT-5 scores 35.

2 and Grok-4 scores 38.6), while missing ones were re-checked under the same conditions and marked with an asterisk. For IMO-AnswerBench, Moonshot re-ran GPT-5 through the API and got 76, versus 65.

6 in the benchmark paper. Coding results are averaged over five runs on Moonshot's own harness, derived from SWE-agent, with trimmed context windows for the Bash and Edit tools and a rewritten system prompt. An independent, reproducible run on a different harness may produce different numbers.

Who can use it, and how

The model is worth trying where a long chain of actions matters: research search across tangled information, multi-step terminal tasks, and frontend work in HTML and React, where Moonshot claims a notable gain. Teams building their own agents should build context management into the pipeline. In its tests, when the 256k window overflowed, Moonshot simply hid all earlier tool outputs, and this trick is easy to replicate.

To evaluate the model on your own data, run it with the full set of tools and step limits close to the benchmark ones: the kimi.ai chat is not suitable for this, while the API is closer to the test conditions. The INT4 gain matters most to those who pay for latency and GPU memory: generation that is twice as fast noticeably changes the economics of agents with hundreds of steps.