Two screenshots show the same project dashboard. The first was taken before a change to the completion filter. The second was taken after. Both show a completion rate of 84%. If an assistant correctly reads the number but applies it to the wrong population, visual accuracy has still produced a wrong answer.
Tencent’s November 4 Hy4 announcement adds visual understanding alongside improvements Tencent reports in long-context and long-horizon work. It discusses Hy4 in hosted products including TokenHub, WorkBuddy, and CodeBuddy. For an enterprise team, the consequential question is whether seeing more representations of a task helps the agent preserve what each representation actually establishes.
Twice launched, with different artifacts
The August 28 preview release shipped weights. Its official repository describes 770 billion total and 49 billion active parameters for that preview backbone, and lists the regular and FP8 preview checkpoints. These numbers and artifacts describe the preview. They are not measurements of the final vision-enabled model’s hardware footprint.
The November announcement describes final Hy4 access in products and says final weights and system cards would follow. On November 14, Tencent’s Hugging Face inventory filtered for Hy4 returned the two August preview repositories. That check establishes what the public query returned; it does not rule out another distribution channel. I did not find a verified final checkpoint or system card in the sources reviewed. Keep the hosted final model and self-hosted preview distinct in any comparison.
There is an immediate implication for buyers comparing architectures. A self-hosted preview can test deployment mechanics for that artifact, while hosted final Hy4 can test an offering that may include different perception and retrieval. A score difference between them cannot by itself be assigned to weights, quantization, or hosting. Match the exact revision first; where you cannot, label the comparison as one between complete solutions.
Look at what an expert study identified
Tencent describes a blind evaluation organized with Tencent Financial Technology: 213 financial professionals, including 153 external experts, reviewing 579 real-world tasks. The release reports Hy4 at 3.04 and DeepSeek V4.1 Flash 1017 at 2.97, with two other models at 2.68. Blinding and realistic tasks make this more informative than a collection of chosen screenshots. The researchers remain the announcement’s source, and the headline comparison alone does not give the paired uncertainty, rater agreement, or task-level distribution needed to interpret the 0.07 difference.
The same announcement reports a lower hallucination rate than Hy4 Preview in an internal search-related evaluation. That might reflect retrieval, model behavior, or both. An answer derived from more current search results is useful even if the reason is better context, but a hosted product gain should not be treated as a property of an offline checkpoint. Ask which tool responses each candidate saw and which claims were checked against primary records.
An office trial can preserve the advantage of expert review without copying the domain. Give reviewers the actual record that determines whether a proposed project update is correct. The report should include the right number, the right entities, and the effective date and dimensions. When the underlying sources disagree, systematic abstention and precise reconciliation can be more useful than a guess that reads well.
Preserve what was seen and what was verified
In a proposed WorkBuddy-style pilot, use a fictional project portfolio. Give the model a slide, a screenshot, and structured milestone records. The three inputs refer to the same project, but only the structured record is authoritative for current status. Render the slide with yesterday’s values and keep the screenshot’s filter visually legible. The task is to write an update with an explanation and references, without editing the underlying records.
There are separate things to score:
| Layer | Observable acceptance criterion |
|---|---|
| Visual reading | Identify the selected period, population, and number from the correct visual region |
| Entity resolution | Tie the pictured item to the correct project and dataset revision |
| Evidence precedence | Use the authoritative current record and explain any conflicting display |
| Artifact handoff | Deliver a report whose links let the recipient check each material claim |
In a second pass, withhold the structured record or move it behind a permission boundary. Score whether the assistant limits its claim to what the visible artifact supports. A good answer can say what a screenshot shows without claiming that it represents the live state. Perception supports that inference only when a separate source authenticates time and scope.
Study the full handoff under change
For a held-out trial, cluster examples by originating workspace or template. Splitting alternate renderings of the same project across development and test can make generalization appear easier than it is. Compare text-only with visual-plus-text, then freeze the model and compare the same inputs with and without an authoritative-record tool. That separates possible improvements in reading from the value of verification. Add a complete product comparison separately, including setup, data access, and review time.
Measure wrong-entity reports, ungrounded equivalence between sources, safe abstentions, and reviewer minutes per accepted report. Add a source revision while a report is being prepared. A useful agent can state its snapshot and rerun the affected checks. If a future workflow allows changes to project systems, the execution contract provides a separate dispatch boundary; no chart, retrieval result, or model-written statement grants that authority.
Hy4’s November release is worth examining because the office workload often moves between visual evidence and structured state. The useful capability is crossing that boundary without losing the identity, date, and meaning of the evidence.