Model Drop

Kimi K2 Thinking

November 6, 2025 2 min read

Kimi K2 Thinking, which Moonshot AI released on November 6, 2025, is the open-weights reasoning version of Kimi K2, built to think step by step while calling tools for up to 300 steps in a row.

Model: Kimi K2 Thinking (moonshotai/Kimi-K2-Thinking)
Maker: Moonshot AI
Ship date: November 6, 2025
Pricing: Free (open weights)
Context: 256K tokens
Weights: Open (Modified MIT)
Where you can use it:
- Kimi chat mode, with agent mode to follow
- Moonshot API and Hugging Face weights
When to reach for it:
- Deep research with hundreds of tool calls
- Hard reasoning paired with web search
- Skip it for image input or quick replies
How it compares:
- Ahead of GLM-4.6 and DeepSeek V3.2-Exp
- Trades wins with GPT-5 and Claude Sonnet 4.5
What makes it different:
- Native INT4 via quantization-aware training
- 200 to 300 tool calls without drift
- 1T parameters, 32B active per token
Early reactions: Smol AI's newsletter called it the first open model to claim wins over top closed models on major benchmarks, while Zvi Mowshowitz noted fewer user reports than a true breakthrough would draw.
Read more: Introducing Kimi K2 Thinking and the model card on Hugging Face

The benchmarks

Moonshot ran K2 Thinking in its thinking mode at temperature 1.0 with 256K context and the native INT4 weights, against GPT-5 at high reasoning effort, Claude Sonnet 4.5 with thinking, Grok 4, its own Kimi K2 0905 and DeepSeek V3.2-Exp; most rival scores are quoted from those makers' posts, and the rest (including GPT-5 and Claude Sonnet 4.5 on HLE with tools) are Moonshot's own reruns. HLE rows use the text-only subset with Hugging Face blocked, and coding scores average 5 runs, in Moonshot's in-house harness except Terminal-Bench (Terminus-2).

Moonshot AI's numbers

Humanity's Last Exam (with tools) (text-only expert questions): 44.9%. +3.2 over GPT-5 (41.7%), +12.9 over Claude Sonnet 4.5 (32.0%), +3.9 over Grok 4 (41.0%), +24.6 over DeepSeek V3.2-Exp (20.3%).
Humanity's Last Exam (no tools) (text-only expert questions): 23.9%. −2.4 under GPT-5 (26.3%), +4.1 over Claude Sonnet 4.5 (19.8%), −1.5 under Grok 4 (25.4%).
BrowseComp (agentic web research): 60.2%. +5.3 over GPT-5 (54.9%), +36.1 over Claude Sonnet 4.5 (24.1%), +20.1 over DeepSeek V3.2-Exp (40.1%).
SWE-bench Verified (real GitHub fixes): 71.3%. −3.6 under GPT-5 (74.9%), −5.9 under Claude Sonnet 4.5 (77.2%), +3.5 over DeepSeek V3.2-Exp (67.8%), +2.1 over Kimi K2 0905 (69.2%).
AIME 2025 (math contest, no tools): 94.5%. −0.1 under GPT-5 (94.6%), +7.5 over Claude Sonnet 4.5 (87.0%), +5.2 over DeepSeek V3.2-Exp (89.3%), +2.8 over Grok 4 (91.7%).
LiveCodeBench (coding problems, v6): 83.1%. −3.9 under GPT-5 (87.0%), +19.1 over Claude Sonnet 4.5 (64.0%), +9.0 over DeepSeek V3.2-Exp (74.1%).
Terminal-Bench (command-line agent tasks): 47.1%. +3.3 over GPT-5 (43.8%), −3.9 under Claude Sonnet 4.5 (51.0%), +9.4 over DeepSeek V3.2-Exp (37.7%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 22.0. +3.4 over MiniMax-M2 (18.6), +3.5 over GLM-4.6 (18.5), +5.4 over DeepSeek V3.2-Exp (16.6), −1.0 under GPT-5 (high) (23.0).
τ²-bench Telecom (agent tool use, Artificial Analysis run): 93.0%. +6.2 over MiniMax-M2 (86.8%), +8.2 over GPT-5 (high) (84.8%), +59.1 over DeepSeek V3.2-Exp (33.9%).