Moonshot AI's Kimi K3.1 Pro Goes 3 Trillion Parameters Open-Weight, Scores 65.7 Percent on Terminal-Bench 4.0, Weights Land by Nov 4

Moonshot AI has officially launched Kimi K3.1 Pro, the first model in its new K3.1 series and what the company calls its most capable 3T-class open-weight model to date, built around a 1-million-token context window. It ships alongside Kimi K3.1 Pro Highspeed, a variant that the company says delivers the same intelligence with roughly 2 to 3 times the output speed.

Both models are available immediately through kimi.com, the Kimi app, Kimi Work, Kimi Code and the Kimi API at platform.kimi.com. The full model weights will be published by November 4 — a date the company is committing to in advance, an increasingly common move among Chinese labs racing to keep open-weight releases credible.

Architecturally, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. Moonshot says it rebuilt every layer of the stack to enable large-scale joint reinforcement learning across hundreds of downstream domains: the RL framework itself, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks. The pitch is that each increment of RL compute now delivers stable, predictable returns rather than noisy, domain-specific gains.

The headline number is Terminal-Bench 4.0, which tests coding agents on realistic, resource-constrained terminal tasks with strict task-specific tests. Moonshot reports K3.1 Pro at 65.7%, ahead of GPT-6 Astra's 58.2% and narrowly behind Claude Opus 5.5's 66.4%. On Kimi Code Bench 3.0 Long-horizon, the company's internal benchmark for multi-step execution, tool use and end-to-end problem solving, K3.1 Pro scores 61.2% against K3's 55.95%. All of these figures are company-reported and have not been independently verified.

Beyond coding, Moonshot says the model shows marked gains in 3D creation, financial research and other long-horizon agentic work — the areas where a million-token context window matters most, since an agent can hold an entire codebase, a full financial filing set, or days of task history in working memory at once.

The launch closes a loop that started in late September, when a kimi-k3-1 identifier first surfaced in Moonshot's API registry and was callable through interface probes before any announcement. Four weeks later, the model is real, benchmarked, and on a published path to open weights.

The competitive frame is unforgiving: OpenAI, Anthropic and Google all refreshed their frontier lines in October, and Chinese rivals MiniMax and Z.ai shipped agentic open-weight models at aggressive per-token prices. A 3T-parameter open-weight release with a committed weights date is Moonshot's answer — betting that enterprises that cannot send data to closed APIs will standardize on the strongest weights they can actually download.

Whether the benchmark claims survive independent reproduction will become clear once the weights land. If they do, the November 4 release could be the strongest open-weight coding model available anywhere; if they drift under outside testing, the gap between vendor-reported and real-world numbers becomes the story instead.