Kimi K2 Thinking was Moonshot's November 2025 reasoning release: a trillion-parameter MoE trained for long chains of thought and sustained tool use, reportedly executing 200–300 sequential tool calls without human intervention. At release it beat proprietary frontiers on several agentic benchmarks — a genuine 'open-weight catches up' moment.
It ships INT4-native thanks to quantization-aware training, roughly halving the memory footprint typical of its class. K2.5 has since superseded it as Moonshot's flagship, but K2 Thinking remains the text-only reasoning specialist in the K2 line and a common choice on hosted inference platforms.
K2 Thinking specifications
| Vendor | Moonshot AI (China) |
|---|---|
| Architecture | Mixture-of-Experts reasoning model, native INT4 (quantization-aware training) |
| Total parameters | 1T |
| Active parameters | 32B |
| Context window | 256K tokens |
| Modalities | text |
| License | Modified MIT — free commercial use; attribution required above 100M monthly active users or $20M monthly revenue |
| Release date | 2025-11-06 |
| Weights | Hugging Face · GitHub |
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| Humanity's Last Exam (with tools) | 44.9% | at release, November 2025 |
| Sequential tool calls without drift | 200–300 | vendor-reported agentic stability |
Running K2 Thinking locally
Ships INT4-native from quantization-aware training (~594GB), which halves memory versus FP8 trillion-parameter peers. Still multi-GPU server territory — but notably efficient for its class.
What K2 Thinking is best for
- Long-horizon agentic tasks and heavy tool use
- Text-only reasoning workloads where K2.5's multimodality is unneeded
- Memory-efficient serving of a trillion-parameter model (INT4-native)
Frequently asked questions
Should I use Kimi K2 Thinking or Kimi K2.5?
K2.5 is the newer flagship and adds native image/video understanding plus Agent Swarm. K2 Thinking remains a strong text-only reasoning model and its INT4-native weights are cheaper to serve. New projects generally start with K2.5.
What does INT4-native mean for K2 Thinking?
The model was trained with quantization-aware training so the released INT4 weights (~594GB) are the intended serving format, not an afterthought — you get the memory savings without the usual post-hoc quantization quality loss.