Kimi K2 Thinking

• 🧠 Moonshot's reasoning agent: interleaves chain-of-thought with live tool calls
• 🏢 Mixture-of-Experts with 1T total and 32B activated parameters
• 📏 256K-token context window, evaluated at full length by Moonshot
• 🔧 Holds coherent goals across 200–300 consecutive tool invocations
• ⚡ Native INT4 weights via quantization-aware training, about 2x faster generation
• 🎯 Moonshot's model card reports 44.9% on Humanity's Last Exam with tools
• 📚 Open weights published on Hugging Face for self-hosting
• 💬 Optional "Heavy Mode" aggregates eight parallel trajectories into one answer

Moonshot

Moonshot is an AI research lab known for developing the Kimi family of large language models. The organization has gained recognition for building capable reasoning-oriented models, with the Kimi line representing its flagship series of text generation…

Kimi K2 Thinking is Moonshot AI's reasoning-focused member of the Kimi K2 line: a large Mixture-of-Experts language model with one trillion total parameters and roughly 32 billion activated per token, trained to interleave step-by-step reasoning with function calls rather than thinking first and acting later. Where the earlier non-thinking K2 releases behaved as instruction-following chat models, this variant is positioned as a "thinking agent," and Moonshot notes it maintains coherent, goal-directed behaviour across 200–300 consecutive tool invocations, compared with prior models that the company says degrade after 30–50 steps.

Two engineering changes define the generational step. The context window is 256K tokens, and Moonshot states its benchmarks were run at that full length. Second, quantization-aware training is applied during post-training with INT4 weight-only quantization on the MoE blocks, giving native INT4 inference and roughly a 2x generation speed-up in low-latency mode without the quality loss typical of post-hoc quantization. On the figures published in Moonshot's own model card, the model reaches 44.9% on Humanity's Last Exam with tools and 60.2% on BrowseComp. An optional "Heavy Mode" rolls out eight trajectories in parallel and reflectively aggregates them into a single answer.

Within the wider Kimi catalogue, K2 Thinking sits between the K2 series and later releases: Kimi K2.5 and Kimi K2.6 continue the general-purpose line, Kimi K2.7 Code targets coding work, and Kimi K3 with Kimi K3 Fast follow as the next flagship generation. Open weights on Hugging Face make K2 Thinking the self-hostable option in that progression.