Kimi K2 Thinking
1T-parameter open agentic reasoning model
About
Kimi K2 Thinking is Moonshot AI's reasoning follow-up to Kimi K2: a trillion-parameter mixture-of-experts model with 32B active parameters, trained end-to-end for agentic work with tool use. Moonshot reports standout agentic numbers — 60.2% on BrowseComp and 71.3% on SWE-Bench Verified with tools enabled — and independent trackers have ranked it near the top of open-weight models. It is open-weight, which is remarkable at this scale, but "open" does not mean cheap to run: serving a 1T-parameter MoE needs serious infrastructure, and most users will meet it through Moonshot's API rather than self-hosting. Benchmark claims come from Moonshot and third-party aggregators and vary by evaluator.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Moonshot/aggregator-reported: BrowseComp 60.2, SWE-Bench Verified 71.3 (tools on), Humanity's Last Exam 44.9, GPQA Diamond 85.7
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Moonshot AI release Nov 2025 confirmed via VentureBeat and aggregators: 1T total / 32B active MoE, agentic reasoning focus.; Benchmarks Moonshot/aggregator-reported: BrowseComp 60.2, SWE-Bench Verified 71.3, HLE 44.9, GPQA Diamond 85.7 — vary by evaluator, flagged.; Open weights per reporting; exact day of release not pinned (releasedAt null).; Limitation: 1T MoE needs serious infra; most users via API; metrics tier (press/aggregator sources, no paper read)..