Overview
- Moonshot published Kimi K3.1 Pro as a long‑context model with a native one‑million token window and a Highspeed variant that it says responds two to three times faster, and the model is available through Kimi.ai, Kimi Work, Kimi Code and the company API.
- Moonshot reported competitive benchmark gains, citing 65.7% on Terminal Bench 4.0 that it says approaches Anthropic’s Claude Opus 5.5 at 66.4% and outpaces OpenAI’s GPT‑6 Astra at 58.2%, and improved long‑horizon coding from 55.95% to 61.2% on its Kimi Code Bench.
- The firm disclosed that it removed cyber‑related training data and tasks during development, which lowered its internal vulnerability score to 81% from 90.5% for an unredacted version, a choice Moonshot frames as a deliberate safety tradeoff.
- Technical claims published with the launch describe a large architecture (reported ~2.8 trillion parameters), hybrid Kimi Delta Attention and a NoPE positional scheme to support the long context and runtime efficiency, but many specifics await independent verification after the weights release.
- Releasing full model weights for public use will let developers self‑host and test the model offline, a move that could widen access and independent evaluation while raising questions about dual‑use risks and cybersecurity oversight.