Model Drop

Kimi K2

July 11, 2025 2 min read

Kimi K2, which Moonshot AI released on July 11, 2025, is a trillion-parameter open-weights model built for coding and agent work rather than long step-by-step thinking.

Model: Kimi K2 (kimi-k2-0711-preview)
Maker: Moonshot AI
Ship date: July 11, 2025
Pricing: $0.60 / $2.50 per M tokens (weights free)
Context: 128K tokens
Where you can use it:
- Kimi on web and app, free
- Moonshot API and Hugging Face weights
When to reach for it:
- Agentic coding and tool use on a budget
- Skip it for long step-by-step reasoning
How it compares:
- Ahead of DeepSeek V3 0324 and Qwen3 235B
- A step above Llama 4 Maverick
What makes it different:
- 1T parameters, 32B active per token
- Trained with the MuonClip optimizer
Early reactions: Nathan Lambert called it the new best open model "by a clear margin," with the modified MIT license as the main gripe.
Read more: Kimi K2: Open Agentic Intelligence and the model card on GitHub

The benchmarks

Moonshot tested Kimi K2 Instruct without long thinking against other non-thinking models (Claude Sonnet 4 and Claude Opus 4 without extended thinking, GPT-4.1, DeepSeek V3 0324, Qwen3 235B in non-thinking mode) with an 8K output cap; some rival scores, like Claude Sonnet 4's SWE-bench number, come from the rivals' own reports.

Moonshot AI's numbers

SWE-bench Verified (real GitHub fixes, one try): 65.8%. +27.0 over DeepSeek V3 0324 (38.8%), +11.2 over GPT-4.1 (54.6%), −6.9 under Claude Sonnet 4 (72.7%).
LiveCodeBench (coding problems, v6): 53.7%. +6.8 over DeepSeek V3 0324 (46.9%), +6.3 over Claude Opus 4 (47.4%), +9.0 over GPT-4.1 (44.7%).
τ²-bench Telecom (agent tool use): 65.8%. +33.3 over DeepSeek V3 0324 (32.5%), +8.8 over Claude Opus 4 (57.0%), +27.2 over GPT-4.1 (38.6%).
τ²-bench Retail (agent tool use): 70.6%. +1.5 over DeepSeek V3 0324 (69.1%), −11.2 under Claude Opus 4 (81.8%), −4.2 under GPT-4.1 (74.8%).
AIME 2025 (math contest): 49.5%. +2.8 over DeepSeek V3 0324 (46.7%), +15.6 over Claude Opus 4 (33.9%), +12.5 over GPT-4.1 (37.0%).
GPQA Diamond (grad-level science): 75.1%. +6.7 over DeepSeek V3 0324 (68.4%), +0.2 over Claude Opus 4 (74.9%), +8.8 over GPT-4.1 (66.3%).
Aider Polyglot (multi-language code edits): 60.0%. +4.9 over DeepSeek V3 0324 (55.1%), −10.7 under Claude Opus 4 (70.7%), +7.6 over GPT-4.1 (52.4%).
SimpleQA (factual recall): 31.0%. +3.3 over DeepSeek V3 0324 (27.7%), +8.2 over Claude Opus 4 (22.8%), −11.3 under GPT-4.1 (42.3%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 12.7. +3.0 over DeepSeek V3 0324 (9.7), +4.4 over Qwen3 235B (8.3), +2.7 over Llama 4 Maverick (10.0), −0.4 under DeepSeek R1 0528 (13.1).