Tencent Hy4 Goes GA With Vision and Computer Use — and No Benchmark Table

Tencent Hy4 became generally available on 4 November 2026, and it is the first model in the Hunyuan line that can read an image and drive a desktop. What Tencent did not ship alongside it is evidence of the kind that sold Tencent Hy4 Preview in August: no Terminal-Bench row, no SWE-bench row, no model card. The release leads instead with internal task suites, a 213-expert finance panel and a hallucination-rate comparison. The most capable Hy model to date is also, on launch day, the least documented one.

That is not an argument for ignoring it. The preview's published scores are still the best available proxy for the family, Tencent says the GA build beats them on the workloads it trained for, and the price did not move. But it does change how you should read the announcement: as a capability claim with a shipping date, not as a measured release. This piece separates the three things Tencent actually shipped — vision, computer use, and a cheaper chain of thought — from the three it has promised and not delivered.

What is actually new in the GA build

Tencent describes three additions, and only the first two are visible to a user on day one.

Vision. Hy4 now handles fine-grained image processing, complex chart analysis and mathematical reasoning over diagrams, combining perception, reasoning and tool calls. Tencent's demo is a 2.8-gigapixel aerial photograph in which the model locates a two-metre roof component — a task that is as much a context-window demonstration as a vision one, since the image has to be tiled and tracked without losing the query. Treat the demo as a vendor illustration; there is no published visual benchmark attached to it.

Computer use. The model can operate a screen through mouse and keyboard and can mix a command line with a graphical interface. Tencent says it trained specifically for Blender, CAD and cross-application long-horizon tasks, and that it runs in a sandbox built on Tencent Cloud's Agent Runtime with several hundred applications installed. This is the more consequential addition, because it is the first time the Hy line claims agency over a desktop rather than over a tool schema.

A shorter chain of thought. The preview shipped with two acknowledged defects: it reasoned longer than necessary on hard tasks and over-verified its own work. Tencent says the GA build improves chain-of-thought efficiency and instruction following, along with long-document understanding and deep research. That claim is testable in a way the vision demo is not — token counts per task are measurable — but Tencent published no before-and-after token figures.

The body underneath is unchanged, at least on paper. The GA announcement does not restate parameters or context length; the figures in circulation are the preview's, 770B total parameters with 49B activated per token and a 1,048,576-token context window. Since Tencent has not published a new architecture description, the reasonable reading is that this is the same backbone with substantially more pre-training and post-training on top, plus whatever vision encoder the multimodal path required — a component Tencent has not described.

The numbers Tencent did publish, and what they are worth

None of these are independent, and none are standardised public benchmarks. They are Tencent's own task suites and expert panels, which makes them useful for direction and weak for ranking.

• Game development — Tencent reports +19.1% on its GameDev suite and +9.8% on GameLogic against the preview. Both are internal suites with no published item counts.

• Finance, expert panel — 213 experts, 153 of them from outside Tencent, roughly 70% with five or more years of experience, rating 579 real financial tasks. Hy4 averaged 3.04 against Kimi K3.1 Pro at 2.68, GLM-5.3 at 2.68 and DeepSeek V4.1 Flash 1017 at 2.97. Tencent concludes it is first among open models on this panel.

• Hallucination — a stated fall from 15.76% to 5.45%, with guess-related errors down from 23.1% to 5.6%. Methodology is not published; treat the size of the improvement as a claim rather than a measurement.

• Research workloads — 讯兔科技's IRAB evaluation is quoted at +12.03 percentage points over the preview, and Tencent's own BioExperimentBench score is 30.4 out of 100. That last number is worth pausing on: it is the one absolute score in the release that is not flattering, and it suggests the biology work is early.

• Training efficiency — Tencent says the model participated in optimising its own training pipeline, lifting end-to-end throughput 20.1%. This is a statement about Tencent's cluster, not about the model you would call.

The absence of the August table is the notable part. When the preview shipped, Tencent published Terminal-Bench 2.1 at 85.4, DeepSWE at 64.3, SWE-bench Pro at 65.7, Toolathlon-Verified at 74.1 and APEX-Agents at 37.1 — the rows that made it credible to people who do not trust blind-panel scores. The GA post does not update any of them. Whether that is because the numbers improved by less than the marketing needs, or because the company is saving them for the system card it has promised, is unknown. Until one of those appears, the honest position is that the GA build has no comparable public evidence.

The price did not move, and the discount is bigger than the price

Tencent kept the preview's rate card. On the Singapore endpoint that is $0.834 per million input tokens, $2.501 per million output and $0.042 per million cache hits; on the mainland card it is ¥6, ¥18 and ¥0.3. A more capable model at an unchanged price is the right way to ship a successor, and it removes price from the upgrade decision entirely.

The mainland announcement adds a promotional stack on top, and the terms matter because they expire at different times:

• Standard API calls at 50% through 31 December 2026, which puts the effective list at roughly $0.417 input and $1.25 output while it lasts.

• Token Plan subscriptions at 2.5折 through 18 November 2026.

• Free quota inside WorkBuddy and CodeBuddy through 12 November 2026.

Those are promotional rates with end dates, not the list price, and they are announced for the mainland market. Budget against $0.834 and $2.501 and treat the discount as upside. The cache rate is the one that compounds: at $0.042 per million, re-reading a large document on every turn of an agent loop is cheaper on Hy4 than on Tencent's own Hy3, which charges $0.059. A model that costs four and a half times more per fresh input token costs less per cached one.

The weights are still a promise

This is the part that decides whether the release matters to anyone who does not want to rent inference.

Tencent Hy4 Preview shipped under Apache 2.0 with full weights, an FP8 variant and ready-made vLLM and SGLang recipes. Anyone can download it today. For the GA build, Tencent says only that it plans to open the model weights and publish a system card with fuller safety evaluations, with ecosystem support from the vLLM and SGLang communities and Inferact. No date, no licence, no repository. A check of Tencent's Hugging Face organisation on launch day finds the preview and its FP8 checkpoint, and no GA repository.

That is a real limitation rather than a formality. A 770B-parameter mixture-of-experts model is not something most teams self-host anyway — the FP8 checkpoint is north of 700GB, and serving it takes a multi-node cluster. But the licence is what decides whether a company can fine-tune, distil, or run the model inside a regulated environment, and until it exists the GA build is API-only. Tencent has also flagged Hy4 Flash as coming soon, aimed at cost and speed, which is the model most teams would actually want for high-volume work.

What you can run today

OrcaRouter's catalogue carries tencent/hy3 and tencent/hy4-preview, including a rate-limited free tier of the preview, and the GA build is not listed at the time of writing. That is worth saying plainly rather than implying otherwise: the model this article is about is not yet routable here.

What the preview does give you is a way to make the decision before the GA arrives. The two share a backbone and a price, so a workload that fails on the preview for reasons of vision or desktop control will be the one that needs the GA; a workload that fails on reasoning quality is the one where the GA's efficiency claims are worth testing. Run the preview against a task you did last week, measure tokens per completed task rather than tokens per call, and you will have a better estimate of the GA's value than the finance panel provides.

When the GA lands in the catalogue it will be a model-string change on the same key rather than a second contract, and a weighted slice with automatic failover back to the preview is the cheap way to try it without betting a production path on a build that shipped this morning. The pass-through pricing model means Tencent's list price is what you pay, so the promotional rate, when it applies, applies here too.

What to watch

Three things will settle the questions this launch left open. The first is a weights date and a licence — that is what separates an open model from an open-washed one. The second is an independent evaluation: every number in this article comes from Tencent, and the preview's eight-month-old table is being used as a proxy because nothing else exists. The third is Hy4 Flash, because a cheaper, faster sibling at the same capability would change the routing maths more than the flagship does.

Until then, the correct description of Tencent Hy4 is narrow and accurate: a generally available API model with vision and computer use, priced identically to its preview, sold on vendor evidence, and waiting on the paperwork that would make it open.