Chinese AI firm Moonshot AI has unveiled Kimi K3.1 Pro, an open-weight model that it says it deliberately weakened in one area: finding and exploiting software vulnerabilities. According to Reuters, the company removed cyber-related data from the training material and reinforcement-learning environments used to build the model, saying it wanted to avoid pushing an openly released system to the frontier of cyber capability. The standard version is available now through Kimi's website, app, coding tools and API, with the full model weights due for release on November 4.
Moonshot put a number on the trade-off. On its internal cyber benchmark, which involves finding and exploiting software flaws in a controlled setting, K3.1 Pro scored 81 percent, against 90.5 percent for an otherwise comparable version built without the risk-reduction measures. The company has not had those figures independently verified. The restraint is notable because it is self-imposed and, for an open-weight model, close to irreversible: once weights are published, the original developer loses control over how others run or modify the system.
On capability, Moonshot is pitching K3.1 Pro as a generalist. It ships with a one-million-token context window and what the company describes as improved coding and reasoning, and it can combine code, web search and 3D generation to build interactive software. In one demonstration it produced a battle game featuring a floating palace, a warrior and a dragon. The model can also generate 3D objects from text, sketches or reference images, feeding a new "3D Make" tool for previewing designs, exporting files for 3D printing, or producing videos. A separate financial-research mode targets models, reports and presentations, and a "Highspeed" variant runs two to three times faster.
The timing is pointed. Moonshot's decision to dial back cyber-offensive ability lands in the same week that researchers warned AI-discovered flaws were already knocking government websites offline, and that the number of vulnerabilities found had doubled year on year. Governments and security experts have grown increasingly uneasy that code-writing models can automate parts of a cyberattack, and the gap Moonshot disclosed, roughly nine percentage points, is a rare public estimate of how much offensive capability a developer is choosing to leave on the table.
It also sharpens the open-weight dilemma. Open models let developers inspect, adapt and run systems locally, advantages Moonshot clearly wants to keep. But safeguards trained into weights can, in principle, be fine-tuned away by whoever downloads them, so the real test of this release is not the launch-day benchmark but whether the reduced capability survives in the versions that circulate after November 4. That it comes from a Chinese lab is significant too: much of the public debate over dangerous model capabilities has centered on US developers, and Moonshot is now setting a disclosed, if unverified, marker of its own.