Overview
- Kimi announced K3.1 Pro and a K3.1 Pro Highspeed variant on Saturday, Oct. 24, 2026, and made both models available across kimi.com, the Kimi apps and the platform.kimi.com API.
- Company-published benchmarks claim K3.1 Pro matches or exceeds leading closed models on coding, 3D and finance tests, and report a drop in targeted factual‑error rates from 23.4% to 8.3% in an internal pressure test.
- Kimi published per‑million‑token API pricing and a Highspeed tier that trades roughly 2–3× throughput for about 2× the price, with cached and uncached input and output rates shown on the platform.
- Developer traces and open‑source commits that surfaced before the launch, including SGLang edits and Hugging Face references, show tokenizer/encoder compatibility work and hint at a new medium inference mode between low and high.
- Kimi says it will publish full model weights by Nov. 4, 2026, which will allow independent researchers and tool builders to verify performance, adapt the model for developer and 3D workflows, and integrate it into third‑party products.