大模型网摘Kimi K3.1 Pro After One Week: Benchmark Readings, Kimi Code Requirements and Open Questions
It has been a week since Kimi K3.1 Pro launched (October 23). Rather than running our own tests, this review assembles the official tech blog, Kimi Code documentation and published benchmark readings to answer three questions: how to read the numbers, what hard requirements apply in practice, and what remains open. See our release coverage for the launch details.
Benchmark Readings: At the Edge of the Leading Group
The most telling figure comes from Terminal-Bench 4.0: K3.1 Pro scores 65.7%, just under Claude Opus 5.5's 66.4% and above GPT-6 Astra's 58.2%. On Kimi Code Bench 3.0 Long-horizon, which tests multi-step execution, the official number is 61.2%, a clear step up from K3's 55.95%; BenchCAD (3D) reads 92.7%. The latter two are self-reported and await third-party verification, so read them separately from Terminal-Bench — at least the latter has same-period competitor numbers attached.
Three Hard Requirements Inside Kimi Code
The docs contain several easy-to-miss constraints. First, K3.1 Pro supports thinking mode only — requests with thinking disabled are rejected outright; reasoning_effort offers low/high/max with high as default. Second, Moonshot advises against switching to K3.1 Pro inside an existing K3-series, K2.8 Preview or K2.7 Code Highspeed session, which degrades quality and reduces cache hits — start a new session instead. Third, protocol-translation layers can distort model behavior, so third-party tools should use the Kimi Coding Plan's native Responses, Chat Completion or Message APIs. The model comes as k3.1-pro and k3.1-pro-highspeed — the Highspeed variant runs roughly 2-3x faster at about 2x quota usage. Availability starts at the Plus plan, with 1M context reserved for Pro and above.
3D and Long-Horizon: What the Demos Show
The official demos emphasize 3D creation and spatial agents: working with Unity, Unreal and Godot, plus interactive Three.js scenes. The richest case is a walking simulator built on the open-source human-anatomy model human-atlas: across 167 tasks and 420 independent reviews, the model probed its own simulator, located and fixed 36 errors in shoulder muscle moment-arm calculations, and progressed from falling within seconds to 11.2 seconds of continuous alternating gait. One demo proves little on its own, but the self-directed error-hunting it showcases is exactly what long-horizon agent benchmarks try to measure.
Open Questions
One: full weights are promised by November 4, but license and quantization details are undisclosed — private-deployment cost modeling has to wait. Two: the memory and bandwidth demands of 3T-class parameters on local or hosted inference remain untested until the weights land. Three: real-world cost against DeepSeek's Flash tier — our V4.1-Flash-1017 vs K3.1 Pro piece works through the metered-vs-subscription math, and we'll revisit deployment once weights ship. For coding-agent integration status, see our coding agent roundup, and for the open-weight field at large, the October flagship spec sheet. Overall: consistent benchmark and documentation posture, clear constraints — a strong open candidate for coding-agent rotation, but hold off on private-deployment decisions until the weights and license land.