Moonshot releases Kimi K2 Thinking, an open model that interleaves reasoning with tool calls
Moonshot's K2 Thinking (1T total parameters, 32B active, 256K context, native INT4 quantisation-aware training, modified-MIT licence) interleaves reasoning with tool calls and is reported to sustain…
- Date
- 6 November 2025
- Who
- Moonshot AI
- Confidence
- High
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): Moonshot AI · Confidence: High Moonshot's K2 Thinking (1T total parameters, 32B active, 256K context, native INT4 quantisation-aware training, modified-MIT licence) interleaves reasoning with tool calls and is reported to sustain 200 to 300 sequential tool calls; Moonshot reported HLE with tools 44.9%, BrowseComp 60.2% and SWE-bench Verified 71.3% (Moonshot page, Willison's launch note). It came 290 days after R1 from the lab that had published the parallel k1.5 recipe, and it is the open-weight example of the "reasoning plus tool use" convergence (B08-24). In 2026 Anthropic alleged Moonshot's accounts had harvested Claude traces (B08-14), and Moonshot's K3 followed.