AI-generated illustration by ByteBriefly, November 4, 2026: a visual-reasoning metaphor. It is not a Hy4 interface, architecture diagram or performance result.
Short answer
Tencent’s Hy4 formal release brings a concrete new model story to the query. November 4 reporting describes added visual understanding and computer-use capabilities. Tencent Cloud’s own documentation now lists the model identifier hy4, image-understanding examples and a 1024k context window. The important access distinction is that an available API and downloadable model weights are separate things; the earlier Hy4-preview release should not be treated as the final model merely because its files are already public.
What is the new Hy4 release?
Tencent’s August 28 announcement introduced Hy4 preview and its open-source availability. The newer formal release is described in Tencent Cloud’s current model table as improving agent and long-horizon productivity capabilities and adding vision. The documentation also distinguishes hy4 from hy4-preview as separate identifiers. Those official details provide a firmer basis than assuming that a familiar family name means an unchanged model.
IT Home’s November 4 account reports computer operation combining graphical and command-line interaction, with model weights and a system card planned for later publication. The dedicated research page returned an access error during this review, so those release details remain attributed to reporting where not independently supported by the API documentation. We have not run the model or reproduced its performance claims.
Why is it trending now?
A formal release with a new input capability gives readers something specific to investigate: whether the model can handle the images and application workflows relevant to them. That is a plausible reason for attention on November 4. The trend snapshot cannot establish which feature attracted a particular searcher or demonstrate that the model performs reliably on an individual task.
The transition from preview to formal release also raises practical questions about model selection. A developer may encounter both identifiers in an application or provider catalog. Checking the exact identifier, enabled tools and documented input types is more useful than relying on a broad description of the family.
Archived artwork from Tencent’s August 28 Hy4 preview announcement. © Tencent. This labels the earlier preview, which must be distinguished from November’s formal release and its planned weights.
What people are looking for
Does vision mean every image task is reliable? An input capability is not a quality guarantee. Test representative charts, diagrams or photographs and compare the answers with known ground truth. Include difficult or ambiguous examples rather than evaluating only a polished demonstration. We supply no independent accuracy score in this article.
Does computer use make an agent fully autonomous? Tool access, application permissions and the surrounding workflow determine what an agent can actually do. Inspect the proposed action and verify the resulting state, especially when a task changes files or external systems. A capability announcement does not establish safe unattended operation for every environment.
Can I download the final release now? Distinguish the dated preview repository from any later final-model release. Look for the actual files, their model card, license and version before claiming local availability. A statement that weights are planned does not itself provide them, and a provider endpoint is a different access route.
How should teams compare it with the preview? Use the same representative workload, tool configuration and success criteria. A longer context limit is an upper bound, not evidence that every long document is handled correctly. Record corrections, incomplete tasks and latency alongside successful outputs; a single appealing example cannot establish production suitability.
What happens next
We will check on November 7 for the final weights, a system card, official access corrections or a substantive independently documented evaluation. Those would answer different questions from API availability and could update this page. Until then, the grounded starting point is a documented formal-model endpoint with reported new capabilities, evaluated within the limits of the evidence actually available.
Snapshot note: the US Trending now capture at 2026-11-04T12:00:02.810926+00:00 records a 500+ traffic bucket and 75% growth against a predicted baseline over a 24-hour collection window. These are bucketed estimates, not exact searches, unique people or a live counter. Started trending (2026-11-04T06:10:00+00:00) is not the event date. Explore’s normalized 0–100 index is separate. Collector values have not been independently verified.