Open Source · NOVEMBER 7, 2026
Tencent’s Hy4 adds computer use; full weights remain promised
Vision and desktop interaction suggest a customer-follow-up pilot for small teams, but Tencent’s internal scores do not establish reliability.
Tencent released Hy4 on November 4 with visual understanding and computer operation, while promising model weights and system cards "in the coming days" in its launch announcement. That separates the capability announcement from the materials needed to examine it. For a small-business owner, the question isn't simply whether an agent can operate a desktop, but whether it can reliably prepare customer follow-ups.
Tencent says Hy4 can interact with command-line and graphical interfaces, interpret complex charts and perform fine-grained image processing. SiliconANGLE and Caixin Global also describe the addition of vision and computer use. Tencent’s headline workplace results, an average performance score of 83.18 and a 96.15% pass rate, are company-reported. They don't independently establish reliability in a customer-acquisition workflow.
Access and inspectability remain different propositions.
Tencent lists Hy4 through TokenHub, WorkBuddy and other products, with API access being made available through third-party platforms. Its announced standard API rates are approximately $0.90 per million input tokens and $2.70 per million output tokens. Despite the launch’s open-source framing, full weights and system cards remained pending. The label doesn't settle what an outside evaluator can verify.
The qualifications extend beyond release timing. The South China Morning Post reports Tencent’s acknowledgment of gaps against leading closed models on internal tests, including long-horizon coding and advanced visual question answering. Reuters describes defensive cybersecurity training and internal safety assessments. Those reports provide context, not independent validation of customer-facing work.
Tencent also reports that internal evaluations reduced hallucinations from 12.95% in Hy4 Preview to 7.2% in Hy4, while errors from unsupported assumptions fell from 23.1% to 5.6%. The announcement doesn't establish those rates for follow-up drafting. For an owner evaluating the model, that's the boundary between a reported improvement and evidence relevant to the job.
### A bounded customer-follow-up test
The practical application here's an editorial inference, not a demonstrated result. A hypothetical pilot for a small team, whether B2B or consumer-facing, would give Hy4 a customer inquiry, product facts and relevant lead notes, then test whether it can prepare a follow-up draft across existing desktop tools. The owner should measure corrections and time saved against the current process. A person should remain responsible for reviewing and sending anything customer-facing.
"The practical implication: evaluate a reviewable draft before considering autonomous customer contact."
That’s an editorial recommendation, not a Tencent claim. Related reading on computer-use agents and unattended desktop work offers another evaluation frame, not validation of Hy4.
This assessment has a short shelf life, measured in weeks pending publication of the full weights and system card and independent computer-use evaluations. Until then, the useful distinction is between a model’s ability to act and a business’s evidence that its actions are dependable. Access permits a test; it doesn't establish trust.