K2 Thinking led open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → but used a record 140M tokens and ran at ~8 tokens/s on the base endpoint, so cheap per-token pricing and slow speed both shape its practical cost.
Key takeaways · AI-distilled
Moonshot only cites post-training for the gains, and Artificial Analysis reads K2 Thinking, the first reasoning modelA model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.Full definition → in the K2 family, as further evidence that RL post-training drives reasoning and long-horizon tool-calling gains.
Its #2 Agentic Index rank, behind only GPT-5, rests largely on 93% on tau2-Bench Telecom, a customer-service tool-use benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → and the highest score Artificial Analysis had independently measured.
In coding it wins no eval outright (6th on Terminal-Bench Hard, 7th on SciCode, 2nd on LiveCodeBench) but leads or ties every open-weights model, displacing DeepSeek V3.2 atop the open-weights Coding Index.
It scored 22.3% on Humanity's Last Exam without tools, an all-time high for open-weights models at the time, behind only GPT-5 and Grok 4.
Endpoint choice changes the economics: the base endpoint ($0.6/$2.5 per 1M tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →) puts a full index run at $356, while the ~50 tokens/s turbo endpoint ($1.15/$8) costs $1,172, second only to Grok 4. Weights ship natively in INT4, about 594GB.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
