2 models, native protocol passthrough, billed per token. From 30% OFF of list price.
as of 2026-10-25
Moonshot AI's Kimi has two models on RouteMux, K3 and K2.7 Code. Both always think, and both read images.
Kimi K3 (July 16, 2026) is Moonshot's flagship: 2.8T parameters activating 16 of 896 experts, with a 1M-token context, and Moonshot's docs suggest starting with kimi-k3 when unsure. Depth is set with the top-level reasoning_effort field, which takes low, high or max and defaults to max; sampling parameters such as temperature and top_p are fixed, so leave them out.
Kimi K2.7 Code (June 2026) is a dedicated coding model built on K2.6, with 1T total and 32B active parameters, a 256K context, and about 30% fewer thinking tokens than K2.6. It has no reasoning_effort, and tool_choice only accepts auto or none. Keep it for clients that still send the K2.x thinking parameter; new integrations should start on K3.
Prepaid wallet billed on actual usage. No monthly fee, no minimum, credits never expire. The protocol layer is a native passthrough, not a translation shim: request bodies and SSE streaming match the official API, so clients only need a new base URL.
Every model is priced against the provider's published per-million-token rate at a fixed discount. Cache reads and writes use the same ratio, with no surcharge. The exact per-model rate is in the table above and always reflects current pricing.
Passed through. Request bodies, SSE streaming format, cache control and extended-thinking parameters match the upstream provider. Point your existing SDK at our base URL and the rest of your code stays unchanged.
No. The prepaid wallet is charged by actual token usage, with no monthly fee and no minimum spend. Unused balance can be refunded under our Terms of Service.