Kimi K2-0905 doubles context to 256k and reports roughly 95% first-try tool-call success at about $1/$3 per million tokens. That makes it a cheap 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 → option for agent Act-mode work.
Key takeaways · AI-distilled
Cline puts Kimi K2's diff-edit failure rate at 5%, against 4% for Claude Sonnet 4 and 10% for Gemini 2.5 Pro, and presents K2-0905 as building on the tool-calling strength of the July checkpoint.
context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → grows from 131k to 262k tokens; per Cline, Moonshot tuned attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → for long-context use, and Cline claims coherence holds across the full window without the usual degradation.
On Groq the model serves at about 349 tokens per second, with a 2-3 second warmup on first requests and faster follow-up requests in the same session, according to Cline.
Cline recommends running K2-0905 in Act mode to execute a plan written by a separate reasoning modelA model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.Full definition →, citing Moonshot's note that it improved at frontend coding.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
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.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
