大模型网摘Kimi K3.1 Pro Debuts: 3T-Class Open Weights, 1M Context, Full Weights Due by Nov 4
The open-weight camp just gained another heavyweight. On October 23, Moonshot AI released Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed: the former is a 3T-class open-weight model with a 1-million-token context window, while the latter trades roughly 2-3x the speed for more responsive interactions. Both are available today on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with full model weights promised by November 4.
From K3 to K3.1 Pro: a large-scale post-training upgrade
According to the official tech blog, K3.1 Pro is a post-trained version of Kimi K3 with substantially larger post-training compute. To run large-scale joint reinforcement learning across hundreds of downstream domains, the team optimized every layer of the stack — the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks. The official summary: each increment of RL compute now delivers stable, predictable returns.
The scorecard: coding and agentic benchmarks
A few verifiable numbers. On Terminal-Bench 4.0, which evaluates coding agents on realistic terminal tasks, K3.1 Pro scored 65.7% — above GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%, placing all three on roughly the same tier. On Kimi Code Bench 3.0 Long-horizon, which tests multi-step execution, tool use and end-to-end problem solving, K3.1 Pro scored 61.2% versus K3's 55.95%. On the 3D-oriented BenchCAD it reached 92.7%. Note that the last two figures are self-reported; third-party verification will take time.
3D and games: beyond web pages
A highlight of the launch is 3D creation and spatial reasoning: K3.1 Pro works with Unity, Unreal and Godot, and can combine web search with 3D generation into complete experiences, with stronger composition, typography and animation in interactive Three.js scenes. One internal case: building a walking-dynamics simulator on top of human-atlas, an open anatomical model with 296 skeletal and 402 muscle components. Across 167 missions and 420 independent reviews, the model iterated on the controller — and even ran open-loop probes to locate and fix 36 errors in shoulder-muscle moment-arm calculations — taking the virtual human from collapsing within seconds to 11.2 seconds of continuous alternating strides.
What it means for developers
On weights, the tonnage of open models keeps climbing: from Qwen3-Max's 1T parameters earlier this year to K3.1 Pro's 3T class, the bar for local and private deployment will be renegotiated. On the API side, 1M context is becoming table stakes — DeepSeek's two current models, V4.1-Flash-1017 and V4-Pro-0813, offer 1M as well. The agentic coding ladder shifts too: we compared V4.1-Flash-1017 against GPT, Claude, GLM and Kimi recently, and a retest with K3.1 Pro is warranted. Meanwhile, DeepSeek open-sourced its production inference framework DeepInfer last week, and the open-weight toolchain from training to serving is getting complete — see our model lineup overview for the full picture.
As of press time the full K3.1 Pro weights have not been released; we will follow up once the license and quantization details are disclosed.