November 4, 2026 is the day Tencent shipped the official release of Hunyuan Hy4, and it deserves an account of its own rather than being folded into the coverage that came before. This site has already treated the August 28 preview open-sourcing in a dedicated piece (the preview open-source report), walked through running that preview on your own machines in a separate guide (the self-hosting SOP), covered the Buddy One office suite unveiled at the October 29 digital ecosystem conference in a third (the Buddy One report), and mentioned the preview only in passing in a Wuzhen open-source story (the Wuzhen report). Each of those keeps a different subject. The focus here is a single event: the official release on November 4 — which capabilities actually changed, whether the weights are available, how the channels and prices are set, and where the model sits in Tencent's wider release cadence.
The headline and the fine print pull in different directions, and that tension is the whole story. When the preview arrived in August, its weights were published alongside the announcement. The official release is a different kind of delivery: products live across Tencent's own apps, API access through aggregator platforms, and a promise that the model weights and a system card will be open-sourced and published in the near future. As of November 4, the weights are not out. Same Hy4 name, two very different bundles. Reading the second as a repeat of the first — assuming the weights came with the release — would get the facts backwards, and it would also erase the single most consequential difference between the two launches.
What Shipped on November 4, and What Did Not
The cadence makes more sense once the timeline is laid out. According to Tencent, the Hunyuan team rebuilt its infrastructure in February 2026 and has since averaged one major version roughly every two months, settling into a fixed pattern of shipping a preview first and a finalized version afterward. On August 28 the Hy4 preview was released and open-sourced, with a total parameter count of about 770B reported at the time by outlets such as China Economic Net and Tide News; that launch is covered in full in the preview open-source report linked above. At the Global Digital Ecosystem Conference on October 29 and 30, Dowson Tong announced that the official Hy4 release was coming soon, sharing the stage with the Buddy One office suite. The official version then landed on November 4, less than a week after the announcement.
In that rhythm, the preview is the phase where a model is handed to the community to be probed and where real feedback accumulates; the official version is where capabilities are consolidated, the model is wired into product lines, and pricing is fixed. A November 4 internet-sector note from CICC judged the official version to be in the same parameter class as the preview — in other words, this release did not buy its gains by piling on parameters. It bought them through a marked expansion of pre-training and post-training, and through broad improvements in instruction following, long-document understanding, long-horizon tasks, and deep research. Holding the base steady while iterating on capability and engineering is precisely what makes a two-month cycle sustainable.
Which brings us back to what was actually delivered. Three things: products live across Tencent's own surfaces, API access through aggregator channels, and an official commitment to open-source the model weights and publish a system card in the near future. The release is usable and callable right now; the weights are not yet open, and the system card has not appeared. This is not a wording quibble. Whether you can obtain the weights decides whether you are limited to calling an API or can fine-tune, deploy privately, and run offline inference. Anyone who cares about self-hosting and VRAM budgets still has to work from the preview self-hosting SOP for now, and wait for Tencent's follow-up before the official weights can be deployed locally.
Two New Capabilities: Vision and Computer Use
Two additions are new relative to the preview: vision understanding and computer use. Vision understanding is introduced for the first time. In Tencent's framing it covers fine-grained image processing, complex chart analysis, and mathematical visual reasoning; the example offered is locating a two-meter rooftop component inside a 2.8-billion-pixel aerial image. The challenge there is not simply recognizing a picture but retrieving and reasoning over precise positions at extreme resolution. Financial charts, engineering drawings, and satellite imagery — inputs that used to be cropped by hand before being fed to a model — can now be handed over whole. Readers who want a systematic comparison of how different models handle visual input can turn to this site's vision input comparison review.
Computer use is also supported for the first time. The model operates software by combining a command line with a graphical interface, and received targeted training on professional tools such as Blender and CAD as well as on long tasks that span multiple applications. According to Tencent, the Agent Runtime team on Tencent Cloud built sandboxes and deployed several hundred applications to train against. Moving from answering questions to physically working inside software is the step that turns an agent from a conversational interface into something that does a job, and it follows directly from the preview's emphasis on understanding, planning, and verifying long-horizon tasks. The other training-side change is scale: both pre-training and post-training were significantly expanded, which Tencent credits with the composite gains in instruction following, long-context comprehension, long-horizon work, and deep research.
The Benchmarks, and the Caveat Tencent Added Itself
Tencent published a fairly granular set of evaluations, and two cautions belong in front of them: most are self-reported, and they take the Hy4 preview as the baseline. On front-end and game development, Tencent worked with its IEG division to adapt the model deeply to Unreal Engine, Unity, and Godot and to extend it to multiplayer scenarios; it self-reports a 19.1% gain on GameDev and a 9.8% gain on GameLogic, both measured against the preview. On office work, the WorkBuddy evaluation shows an average effectiveness score of 83.18 and a task pass rate of 96.15 percent, with Tencent claiming first place among domestic models in nine of twelve scenarios. These figures sketch a model tuned for writing, reading, and carrying long tasks to completion, but because the evaluations and the baseline are the company's own, they read better as a capability profile than as an objective cross-vendor ranking.
Finance received the most attention. Tencent, working with its FinTech unit FiT, organized a crowd evaluation: 213 experts took part, 153 of them from leading external financial institutions, with roughly seventy percent bringing more than five years of experience, across 579 real financial tasks. Inside that company-run blind test, Hy4 scored 3.04, Kimi K3.1 Pro and GLM-5.3 each scored 2.68, and DeepSeek-V4.1-Flash-1017 scored 2.97. Tencent calls the result the best overall performance among open models — and it is worth being exact about what that means. It is a self-reported conclusion from Tencent's own 579-task blind test, not a cross-leaderboard objective ranking; the task sets, graders, and setups of different evaluations do not line up, so lifting that sentence out as an industry-wide placement would distort it.
The more interesting move is that Tencent supplied both sides of the judgment itself. On one side, it describes Hy4 as entering the ranks of the strongest open models for the first time, which is likewise a self-reported characterization and should be read with that qualifier attached. On the other side, the very same official materials state plainly that the model still has a considerable gap from leading closed models. One sentence locates it inside the open camp; the other measures its distance from the closed frontier. Only together do they give an honest position, and quoting either one alone tilts the picture. Internal research evaluations show the same restraint: TenureTrail spans seven disciplines and 123 executable research tasks, while the Hy-BioExperimentBench biosafety score is 30.4 out of 100, which Tencent itself says leaves a visible gap to frontier models. On safety, cybersecurity training drew only on defensive data such as proof-of-concept generation and vulnerability patching, and was audited by Tencent's Xuanwu and Keen labs. A launch note willing to print strengths and shortcomings side by side earns more trust than one-sided promotion.
Where the Parameter Figures Come From
One more piece of housekeeping: the parameter numbers should not be mistaken for an official specification sheet. The official release page does not publish a standalone parameter table. The 770B total comes from media reporting by outlets such as China Economic Net and Tide News. The 49B active parameters and the 1M context window come from CICC's judgment that the official version sits in the same parameter class as the preview. The precise context length of 1,048,576 is what an OpenRouter capture of the tencent/hy4 entry on November 4 returned as its context_length, where the modality field indicates text and image input producing text output; the same platform's tencent/hy4-preview entry is text to text only. Three separate provenances, each with its own source, together outline the scale — but none of them should be written down as a specification Tencent officially announced for the release. Keeping media figures, analyst judgments, platform captures, and official specifications in separate columns is the basic hygiene of reading a launch like this.
Channels, and Three Price Sets That Do Not Explain Each Other
Channels are worth reading item by item, arguably more than parameters are. The live entry points Tencent lists include TokenHub on Tencent Cloud, WorkBuddy and CodeBuddy, Yuanbao, the Dayuan assistant in WeCom, and ima; QQ, Honor of Kings, Peacekeeper Elite, WeChat Reading, and Tencent News are in the process of being connected. On the API side, access arrived simultaneously on OpenRouter, Hermes, and OpenCode. Readers ready to wire up an API can follow this site's TokenHub API integration SOP.
Pricing has to be read as two unrelated sets. First, the Tencent Cloud list price in yuan, unchanged from the preview: 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million tokens on a cache hit. Around that list price sit several official promotions: through December 31, 2026, TokenHub is billed at fifty percent of the list price; Token Plan subscriptions carry a twenty-five percent credit-consumption rate until November 18; and WorkBuddy and CodeBuddy offer a certain free quota from November 4 to November 15. Note the phrasing — a certain free quota, not unlimited access, and a defined window for hands-on evaluation rather than a blanket giveaway.
The second set is the dollar pricing captured on OpenRouter on November 4, and it stands apart from the Tencent Cloud yuan list price rather than converting or explaining it. On that channel, tencent/hy4 is listed at $0.417 per million input tokens, $1.2505 per million output tokens, and $0.021 per million tokens for cache reads, with a context_length of 1,048,576 and a modality that accepts text and image input and produces text output; for comparison, the tencent/hy4-preview entry on the same platform is text to text, at $0.834 input, $2.501 output, and $0.042 for cache reads. One point deserves emphasis: the TokenHub fifty-percent promotion applies only to the Tencent Cloud list price and does not carry over to the OpenRouter dollar channel. The two channels keep separate pricing, separate promotions, and separate settlement currencies, so using the dollar figures to reverse-engineer the yuan list price, or applying the discount to the dollar figures, will produce the wrong number.
Hy4 Flash, Hy5, and the Release Cadence
The official version is not the end of the line. Tencent also previewed a cost-focused Hy4 Flash for the near future and said Hy5 would scale up further; add the commitment to open-source the model weights and publish a system card in the near future and the follow-up calendar is crowded. Set against the rhythm established in February 2026 — a major version roughly every two months, preview first and finalized afterward — what Tencent is demonstrating is not a single benchmark score but a predictable delivery cadence: the preview opens to the community, the official version consolidates and goes live, a flash variant covers the value end, and the next generation scales up. Four tracks in parallel leave room for community freedom, enterprise stability, and interfaces pointing both toward lower cost and toward higher capability.
For an industry reader, that cadence is more useful than any one leaderboard position. It means selection can be planned around versions rather than re-evaluated from zero each time, and it means the timing of the weight release will directly shape the ecosystem: until the weights are actually out, the community can call APIs, build applications, and accumulate feedback, while fine-tuning and private deployment have to wait. As for the model's standing among open models, the honest answer stays inside Tencent's own two qualifiers: Hy4 is described as entering the ranks of the strongest open models for the first time, which is self-reported, while the same materials acknowledge a considerable gap from leading closed models. What actually landed on November 4, 2026 is a fuller-featured official version with broader distribution and clearer pricing, plus an as-yet-unfulfilled promise to open-source the weights. The word open only earns its place back in the headline once the weights and the system card are truly released.
Related reading