Tencent launches Hy4 flagship with computer use and upcoming open weights

Top-tier AI models have already mastered writing basic code, so the frontier has shifted toward letting models actually see and operate computers.

Tencent just unveiled its new flagship model, Hy4. According to benchmark results, the model delivers performance right alongside the top open models available today.

Hy4 arrives with native vision capabilities and built-in computer use features, positioning it directly for agent-style desktop automation.

Available now, open weights next

Even though Hy4 is destined for open weights, you do not have to wait to start testing it.

The model is already live on Tencent Cloud and OpenRouter. It is also wired directly into platforms like CodeBuddy, WorkBuddy, and Hy AI Studio for early hands-on use.

Tencent confirmed that the open model weights are coming shortly.

Instead of just dropping files into a repo, Tencent partnered early with the vLLM and SGLang communities to guarantee ecosystem and runtime support. When the weights land, Tencent will also release a comprehensive system card detailing safety evaluations and deeper technical benchmarks.

Why it matters

We think Tencent is executing this rollout with smart timing and proper engineering hygiene.

For developers building autonomous workflows, models capable of reliable computer use are in high demand. Having a competitive model with native vision and desktop interaction opens up practical automation tasks that previously required expensive closed APIs.

Tying up with vLLM and SGLang ahead of the weight release is a major win for practitioners. All too often, a promising open model drops, but teams spend weeks hacking together custom inference scripts. Ensuring day-one framework support dramatically lowers the barrier to production.

On the broader industry board, the open-weight frontier has largely been contested by players like DeepSeek and Meta. Tencent stepping in with a flagship that targets computer use raises the stakes for everyone racing to build practical AI agents.

That said, benchmarks are only half the battle. We will see how well Hy4 holds up once the broader community stress-tests its vision-driven computer actions against messy, real-world user interfaces.