Moonshot AI Kimi K3
A 2.8-trillion-parameter mixture-of-experts model — the largest open-weight model shipped to date — with native multimodal understanding, a 1M-token context window, and reasoning always on. Priced at $3 per million input tokens and $15 per million output. Scores 57 on the Artificial Analysis Intelligence Index (4th of 189, ahead of Opus 4.8, behind Fable 5 and GPT-5.6 Sol).
API live July 16 including OpenRouter; full open weights promised by July 27 but not yet published as of this entry.
Reality review pending — we publish the executive verdict after the launch has had time to meet the real world.
An open-weight model now sits inside the top four overall, beating a current Anthropic flagship on independent tests — the strongest evidence yet that self-hostable models can be genuine frontier substitutes. This reframes build-vs-buy: frontier-adjacent capability with an eventual path to running it on your own infrastructure, with the usual China-origin governance questions attached.