Moonshot releases Kimi K2 Thinking, a 1T MoE thinking agent with native INT4
Kimi K2 Thinking is a 1T-parameter MoE with 32B active, 256K context and native INT4 quantisation, under a modified MIT licence.
The moonshotai organisation published Kimi K2 Thinking, which its card calls the latest and most capable open-source thinking model in the line. It is built as a thinking agent that reasons step by step while calling tools. The card claims stable behaviour across 200 to 300 sequential tool calls.
The model is a mixture-of-experts with 1T total and 32B activated parameters, 61 layers, 384 experts with 8 selected per token, and a 256K context. Quantisation-aware training in post-training gives native INT4 weights, which the card says double speed in low-latency mode without loss. The card claims state-of-the-art results on Humanity's Last Exam and BrowseComp. Those are the maker's claims.
Both code and weights are under a Modified MIT License. Kimi K3, covered separately, is the later model.
The card gives no release date; the date used is the Hugging Face repo creation date.
An older sibling now, though still a reference for open tool-using models.
Sources
Editor's first pass (AI-assisted)
- Real 3.5: Ungated weights under a modified MIT licence; 1T parameters even at INT4 needs a large cluster.
- Useful 3.5: Tool-heavy reasoning and research workflows, with 200 to 300 sequential tool calls claimed.
- Notable 4: Over 1,700 likes; an earlier step on the road to Kimi K3.
- Novel 3.5: Quantisation-aware training for lossless INT4 is the distinct feature.
Comments