Moonshot AI just did something almost no AI lab volunteers to do: it made its newest model worse at one job, on purpose.

On Saturday, the Beijing-based company said it had stripped cyber-related data from the training of Kimi K3.1 Pro, its most capable open-weight model, while keeping the general coding skills that made the Kimi line popular with developers. Reuters reported the curbs as Moonshot introduced the model on October 23, and the company plans to publish the model’s weights by November 4.

“We aim not to release open weights that would extend the absolute frontier of cyber capability,” Moonshot said, according to Reuters.

On Moonshot’s internal xdev-bench-hard test, the shipped model scored 81.0%. An unfiltered checkpoint of the same model, trained with the cyber data left in, scored 90.5%. Kimi K3, the previous generation, scored just 33.3%.

Read that plainly: the filtered model is still far more capable at exploitation than its predecessor.

What Moonshot changed in Kimi K3.1 Pro

Moonshot removed cyber-related data from both the training corpus and the reinforcement-learning environments used to build Kimi K3.1 Pro. General coding ability was deliberately left intact, so the model can still write software, use tools, and work through long, multi-step tasks. What it lost was the specialized diet of data that teaches a model to find and exploit software flaws.

Timing explains part of the decision. Governments, tech companies, and security researchers are growing worried that AI could make sophisticated hacking easier and cheaper, since a model that writes code and uses tools on its own can also probe programs for weaknesses.

China added to that backdrop last week with proposed export controls that would ban exports of AI-driven malware-generation and mutation technology and restrict AI-assisted vulnerability-discovery tools. Those draft measures, still open for public comment, cite risks to critical infrastructure, defense-industrial systems, and core government systems.

Kimi K3.1 Pro benchmarks, specs, and price

Kimi K3.1 Pro is the first model in Moonshot’s K3.1 series and a post-trained version of Kimi K3, built with what the company describes as substantially larger post-training compute. It runs a mixture-of-experts design with 2.8 trillion total parameters, 104 billion of them active at any one time, and it handles a 1-million-token context window.

Moonshot’s own numbers put the model close to the closed frontier on coding work. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, Kimi K3.1 Pro scored 65.7%, against 66.4% for Anthropic’s Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. On BenchCAD it scored 92.7% with tools, behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. On Kimi Code Bench 3.0 Long-horizon it scored 61.2%, up from 55.95% for K3.

Those are vendor numbers, and they deserve the usual caution. Independent tester Artificial Analysis gives the model an Intelligence Index score of 55, fourth overall behind only Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56. Artificial Analysis also measured 1,846 Elo on GDPval-AA, 312 points above K3, and a Terminal-Bench 4.0 score of 67%, the highest it has measured. On its own hallucination test, the rate fell to 44% from 53% for K3, while Moonshot’s internal stress test showed claim-level errors dropping from 23.4% to 8.3%.

Price is part of the pitch. Moonshot charges $2 per million input tokens and $10 per million output tokens through its API, and Artificial Analysis estimates about $2.11 per Intelligence Index task, against $5.98 for Opus 5.5 at maximum effort. A Highspeed variant runs two to three times faster for buyers who need lower latency.

Why the November 4 weights release matters

Publishing weights changes who controls the model. Reuters notes that once open-weight files are widely distributed, the original developer has less ability to restrict access, monitor use, or update safeguards than with a hosted service. A hosted model can be patched or pulled. A downloaded one cannot.

Moonshot’s counterargument is that openness helps defenders too. The company says open-weight models let more organizations inspect and secure their own software, including older systems that receive little attention from vendors. That trade is the whole debate in one line: the same openness that lets defenders inspect the model also puts it beyond Moonshot’s control after November 4.

Why this matters

Don’t read the 81.0% figure as a safety guarantee. Moonshot slowed the climb in hacking skill; it did not reverse it, and the filtered model still scores more than twice as high as K3 on the company’s own exploit test. What changed is the direction of travel.

For US developers, the practical question is what actually lands on November 4: a near-frontier coding model they can run themselves, priced well under closed rivals, with a documented limit on one dangerous skill. Watch whether independent testers reproduce Moonshot’s scores once the weights ship, and whether rival labs answer with published limits of their own. If they do, Moonshot’s curb becomes a precedent. If they don’t, it stays a curiosity.

Frequently asked questions

Why did Moonshot weaken Kimi K3.1 Pro’s hacking ability?

Moonshot says it does not want its open-weight release to push cyber-offensive capability further, so it removed cyber-related data from training while keeping general coding skills.

How much did the curbs reduce its hacking score?

On Moonshot’s internal xdev-bench-hard test, the shipped model scored 81.0%, down from 90.5% for an unfiltered version. Kimi K3 scored 33.3%.

When will Kimi K3.1 Pro’s open weights be released?

Moonshot plans to publish the weights by November 4, 2026. Until then, the model is available through the Kimi app, Kimi Code, and the Kimi API.

How does Kimi K3.1 Pro compare with Claude Opus 5.5?

On Moonshot’s Terminal-Bench 4.0 figures, Kimi K3.1 Pro scored 65.7% against 66.4% for Claude Opus 5.5. Artificial Analysis ranked it fourth overall with an Intelligence Index score of 55, behind Opus 5.5 at 58.

Sources: Reuters, Unite.AI, SiliconANGLE, Artificial Analysis