Kimi

Core capabilities

Flagship and Highspeed tiers. The open platform sells K3.1 Pro as Standard and Highspeed: the capability description is the same, Highspeed targets higher concurrency and lower latency, and its unit price is about twice Standard's. For caching, the homepage lists write and hit prices separately; hits are much cheaper than writes (CN hit ¥1.30 / M tokens, international hit $0.20 / M tokens, as of 2026-10-25). The homepage lists only one write price with no TTL; the pricing page and context-caching docs say K3-series cache writes are billed separately by TTL, 5 minutes (default) or 1 hour, with the 1-hour tier priced higher. For kimi-k3 (per M tokens): CN input (miss) ¥20, write 5m ¥20 / 1h ¥40, hit ¥2; international $3, write 5m $3 / 1h $6, hit $0.30. K3.1 Pro supports about 1M context and vision input; parameter counts, benchmarks and capability statements are all vendor claims.

Open weights ≠ open product. Kimi-K3 on Hugging Face can be downloaded under the Kimi K3 License for self-run inference; the consumer apps, web, Kimi Code / Work / Claw and the hosted API remain closed commercial services. The license grants MIT-style rights (use, copy, modify, distribute, sublicense and sell) with several added conditions (the original text governs): section 1, keep the copyright and license notices; section 2, if the licensee or an affiliate runs a MaaS business and the licensee together with its affiliates has total revenue above USD 20M in any consecutive 12 months, a separate agreement with Moonshot is required before any commercial use (the license defines MaaS as an inference or fine-tuning service that lets third parties substantially control inputs, parameters or training data; features merely embedded in an end product, or simply relaying requests to models hosted by others, don't count); section 3, use in a commercial product with more than 100M monthly active users or more than USD 20M monthly revenue requires prominently displaying "Kimi K3" in the user interface; section 4, internal use (not providing the software, its outputs or capabilities to third parties), and use through Moonshot's official products or certified inference partners, are exempt from sections 2 and 3.

Agent and coding line. The membership help center describes the tiers: Go covers documents, slides, deep research and website deployment, without Kimi Code; Plus and up add Kimi Code and Agent clusters; Pro and up emphasise goal mode and a dedicated Kimi Claw; Max / Ultra raise the limits on parallelism, scheduled tasks and project storage (the help center's Ultra example: about 28× Agent quota, 4 parallel tasks, 25 each of scheduled and widget tasks; the purchase page governs). Kimi Code also has a rolling 5-hour usage cap (new plans have no weekly cap; legacy plans such as Andante keep it). The help center also says a deployed Claw cloud host, even with no active task, deducts a sandbox fee of about 0.6% of membership quota per day (settled at 4 p.m. daily), and a website published by an Agent costs about 0.08% for hosting. Charges stop only once the files are saved and the cloud host deleted, or the site unpublished.

Enterprise options. The open-platform docs say the Kimi API open platform is pay-as-you-go with no subscription plans, unlike Kimi membership, Kimi Code and other products. "No subscription plans" refers only to how the API open platform bills, not to Kimi membership or Kimi Code subscriptions; the open-platform docs navigation also has "Enterprise subscription" and "Contact sales" entries, plus a Zero Data Retention (ZDR) page for customers with requirements on inference log retention. Eligibility, covered models and contract text are as set in the console and commercial terms.