Tencent Hy4 ships with cyber-offense filtered out

Tencent released Hy4 on Wednesday, and the headline is not a benchmark. It is what the model was not trained on. Per Reuters, Tencent filtered exploit development, reverse engineering and penetration-testing material out of the training data to limit offensive cyber capability. Weights and a safety report are promised later. If you put the Hy4 preview in your fallback plan in August, this release changes two things about it.

What actually happened

Tencent's statement, as reported by Reuters and Business Day:

• Cybersecurity training content was restricted to defensive uses, such as identifying software flaws and helping fix them. Material on building hacking tools, breaking into systems and testing systems for weaknesses was excluded.
• Internal security specialists tested Hy4 and found no sign of higher risk in helping users break into software or bypass protections.
• Testing found the model not capable of meaningfully easing harmful biological experiments.
• Tencent calls Hy4 open-source and says it plans to release the weights later, with a report describing its safety testing. No date.
• Neither report gives pricing, context length or benchmark numbers. The August preview was a 770B-parameter mixture-of-experts model with 49B active, 1M context, under Apache 2.0.

Context Reuters supplies: Moonshot AI limited its newest model's cyber-offensive capabilities last month, and Tencent disclosed in October that agents it used to test third-party models showed unintended browsing behavior toward Alibaba's Amap service. Tencent said that activity did not come from its Hunyuan models.

Why Hy4's training filter matters for your business

Two operator-level points.

First, "open-source" with weights "later" is a hosted model until the files land. The preview weights are on Hugging Face under Apache 2.0 and nobody can take them back. The new Hy4 is, today, an API. Do not update your exit plan on a promise. Keep the preview pinned as your fallback until the release weights and the safety report are public, then re-run your evals.

Second, capability shaping by training-data exclusion is now a vendor pattern, and it is directional, not surgical. Tencent wants the model to find and fix flaws but not exploit them. Those tasks share a lot of knowledge. If your workflow uses a model to read a CVE advisory, explain a proof-of-concept your scanner flagged, or triage a dependency alert, that is squarely in the zone Tencent filtered. It may work fine. You will not know until you test it on your own tickets.

Our rule: every model that enters a client stack gets the same 30-task eval, and five of those tasks are security-review tasks on our own code. A model that scores well on coding and refuses or hallucinates on "explain why this dependency is flagged" is not a cheaper model. It is a different tool. Vendors shaping capability for good reasons does not change your job, which is to measure the capability you actually depend on.

Sources: Reuters via The Star, Business Day.