Overview
- Moonshot publicly launched Kimi K3.1 Pro on Friday as a 3‑trillion‑class model that natively handles a one‑million‑token context and is available now via the company’s consumer, enterprise and API platforms.
- The company reported agent and coding gains on internal tests, including 65.7% on Terminal Bench 4.0 and 61.2% on its long‑horizon coding benchmark, and said the model scored 92.7% on BenchCAD for 3D tasks.
- Moonshot plans to release the full open weights on November 4, which will let external developers run and inspect the model independently rather than only through hosted services.
- To limit offensive cyber capability the firm removed cyber‑related training data and reinforcement tasks and reports the hardened model scored 81.0% on its internal vulnerability benchmark versus 90.5% for an unredacted version.
- Moonshot also published technical claims about a Kimi Delta Attention hybrid attention scheme, NoPE positional approach, quantization strategies and a faster Highspeed variant that it says runs two to three times quicker, though independent validation will be needed after the weights drop.