Tencent Hy4 vs Tencent Hy4 Preview: The Newer Model Is the Less Open One

Tencent Hy4 Preview has an Apache 2.0 licence, 770 billion parameters you can download from Hugging Face today, and a benchmark appendix with eight rows of published numbers. Tencent Hy4, released on 4 November 2026, has vision, computer use and a shorter chain of thought — and none of that paperwork. Its weights are announced as a plan, its licence is unstated, and its model card does not exist. In the Hunyuan line the version number went up and the openness went down, which makes this less a routine upgrade comparison than a question about which artefact you are actually choosing.

They are the same family, the same size and, as of today, the same price. That combination is unusual enough to be worth stating plainly: Tencent is selling a better model for the same money while giving away less of it. Whether that is a problem depends entirely on whether you intended to run the weights yourself.

Same backbone, one more training run

Tencent has not published a new architecture description for the GA build, and it did not restate the parameter count in the release. The figures that remain in circulation are the preview's: 770B total parameters, 49B activated per token, 78 layers with a dense first layer and 77 mixture-of-experts layers of 256 routed experts plus one shared, top-8 routing, a 1,048,576-token context and a 10B multi-token-prediction layer with 0.7B active.

The GA announcement frames the difference as scale in pre-training and post-training rather than a new design, and describes the result as the largest generation-over-generation gain Tencent has measured. That is consistent with what the company did between HY3 and Hy4 and with how the preview was built. The one component that cannot be carried over unchanged is the vision path — image understanding requires an encoder that a text-only checkpoint does not have — and Tencent has said nothing about its size, its training data or whether it is a separate model stitched on. Until a model card appears, treat "same backbone plus vision" as an inference, not a disclosure.

The three things you can test today

Vision. The GA build reads fine-grained images, complex charts and diagrams used for mathematical reasoning, combining perception with tool calls. Tencent's demonstration is a 2.8-gigapixel aerial photograph in which the model finds a two-metre roof component — partly a vision result, mostly a demonstration that a million-token window can hold a tiled image and a query at once. The preview cannot do this at all; it is text in, text out.

Computer use. The model can drive a desktop through mouse and keyboard and mix a command line with a graphical interface, with specific training for Blender, CAD and cross-application long-horizon tasks. Tencent runs it in a sandbox built on its own Agent Runtime with several hundred applications installed. This is the larger departure from the preview, which was built for tool schemas rather than screens.

Chain-of-thought efficiency. The preview shipped with two documented problems: it reasoned longer than necessary on hard tasks and over-verified its own answers. Both were acknowledged in the model card itself, which is unusually candid and made the preview harder to use in production than its benchmark scores implied. Tencent says the GA build shortens the reasoning and improves instruction following. This is the claim most worth testing, because it is the one that changes tokens per completed task — and Tencent published no before-and-after token counts.

The delta Tencent published

There is no GA benchmark appendix. The comparison Tencent released is against the preview, on its own suites, and it is thinner than the preview's own launch material.

• Game development — GameDev +19.1% and GameLogic +9.8% over the preview, on internal suites with no published item counts.

• Finance — a 213-expert panel, 153 of them external, rating 579 real financial tasks, puts Hy4 at 3.04 against 2.68 for both Kimi K3.1 Pro and GLM-5.3 and 2.97 for DeepSeek V4.1 Flash 1017.

• Hallucination — stated as falling from 15.76% to 5.45%, with guess-related errors from 23.1% to 5.6%. Methodology unpublished.

• Research — 讯兔科技's IRAB evaluation is quoted at +12.03 percentage points over the preview; Tencent's BioExperimentBench score is 30.4 out of 100, which is the least flattering absolute number in the release.

Set that against what the preview published on 28 August: Terminal-Bench 2.1 at 85.4, DeepSWE at 64.3, SWE-bench Pro at 65.7, SWE-bench Multilingual at 82.9, Toolathlon-Verified at 74.1, OneMillionBench with tools at 65.4, SWE Atlas Refactoring at 53.3 and APEX-Agents at 37.1. Those rows are what made the preview legible to anyone outside Tencent. The GA release does not update a single one of them, so there is no way to say whether the GA build is better than the preview at software engineering, only that it is better at game development and finance panels.

What the preview still has that the GA does not

Four things, and each is a practical capability rather than a technicality.

The weights themselves. The preview and its FP8 variant are on Hugging Face, ModelScope, GitCode and CNB under Apache 2.0. The repository was created on 27 August 2026 and had accumulated roughly 11,500 downloads by launch day of the GA build. There is no GA repository.

A licence that permits work. Apache 2.0 allows commercial use, modification, fine-tuning and redistribution with no revenue threshold and no acceptable-use rider beyond the patent grant. Tencent says a system card with fuller safety evaluations is coming for the GA build and that responsible release will cover cyber and biological risk. That language usually precedes a more restrictive licence, and there is no way to know yet.

Serving recipes. The preview ships with vLLM and SGLang deployment instructions and a published set of sampling defaults — temperature 0.9, top_p 1.0, reasoning effort defaulting to high with a no-think mode available. The GA build has named vLLM and SGLang as ecosystem partners without publishing recipes, and vision plus computer use will need serving paths the text-only preview never did.

An auditable benchmark record. Whatever the vendor bias in the preview's appendix, it is a fixed set of rows that anyone can run against the checkpoint and dispute. The GA build has no such record yet.

Which artefact to pick

The price is identical, so cost does not decide it. On the Singapore list both are $0.834 per million input tokens, $2.501 per million output and $0.042 per million cached input; on the mainland card both are ¥6, ¥18 and ¥0.3. Tencent explicitly kept the rate unchanged for the GA release, which is the strongest argument for it: more capability, same bill.

Choose the preview when the weights are the point. On-premises deployment, regulated data, fine-tuning, distillation, or a cost model that assumes you serve it yourself — none of that is available for the GA build, and the preview is a genuinely capable 770B model rather than a crippled teaser. An Apache 2.0 checkpoint that runs on your own cluster is a different asset class from an API endpoint, and no benchmark score closes that gap.

Choose the GA build when the modality is the point. Image understanding, desktop control and long-horizon cross-application work exist nowhere in the preview. If the task involves a screenshot, a chart, a CAD file or a click, the older model is not a cheaper substitute — it cannot do the job at all.

For everyone else — text workloads inside a million-token context, called over an API — the honest answer is that the GA build is probably better and nobody can prove it yet. The efficiency claims are the reason to move early, because a model that stops over-verifying its own output will cut output tokens per task, and that is a cost improvement the rate card does not show. Test it on your own tasks with token counts recorded, not on the finance panel.

FAQ

Will the GA weights be the same checkpoint as the preview?

No. Tencent describes the GA build as a substantially larger pre-training and post-training run, not a re-release of the preview, and a vision-capable model needs components the text-only checkpoint does not contain. What is unknown is how much of the preview survives in it. Tencent has not said whether the GA start point was the preview checkpoint or an earlier one, which matters if you are trying to predict fine-tuning behaviour from published preview results.

Is Tencent Hy4 Preview being retired?

No retirement has been announced. The preview remains on Hugging Face under Apache 2.0 and two OrcaRouter catalogue entries still serve it, including a rate-limited free tier at zero wallet cost. Tencent has positioned the GA build as the production successor at the same price, which in practice tends to strand preview deployments rather than remove them — the weights stay downloadable, the recipes stay published, and the model quietly stops being the one anybody benchmarks.

Does the GA build cost more for images?

Tencent has not published a separate image price, so the published rate is per token across modalities. That is less reassuring than it sounds: image inputs are tokenised, and the number of tokens a chart or screenshot consumes is not disclosed. A dense page image can cost far more than its character count suggests. Measure it on your own inputs before assuming the text rate applies cleanly.

Can I run the GA build locally?

Not today. No weights, no licence, no repository. Even when they land, a 770B mixture-of-experts model at FP8 is north of 700GB and needs a multi-node cluster, so "open weights" here means self-hostable by an organisation with serious infrastructure, not by a workstation. If local execution is the requirement, the preview is the only option in the family and will remain so until Tencent publishes.

The gap that matters is not the score

Tencent's framing is that this is the first Hy model to reach the top tier of open-source models. The claim is defensible on capability and premature on openness, because the version that earned the open-source framing is the one from August, and the version being sold today is an API product with a promise attached.

That is not hypocrisy so much as sequencing. Tencent has said weights and a system card are coming, named vLLM, SGLang and Inferact as partners, and shipped a preview under Apache 2.0 that it did not have to. The pattern from HY3 suggests the release will happen. The pattern also suggests it will take weeks, and that the licence may not be Apache this time — the safety language in the GA announcement is more careful than the preview's was.

In the meantime the two models are both live and both usable, and the choice between them is a choice between an open artefact and a better closed one at the same price. On OrcaRouter, tencent/hy4-preview is callable alongside tencent/hy3 on one key, with the provider's list price passed through and no markup, and a free tier that costs nothing but is rate-limited. That makes the preview the cheapest possible way to evaluate the family before the GA build reaches the catalogue — same weights, same price, one model-string change away, and no commitment while you decide whether the GA's vision and desktop work is worth the missing licence.