Tencent Hy4 adds vision and computer use, tested in WorkBuddy

On 4 November 2026, Tencent released Hy4, the final version of its Hunyuan language model, following a preview published on 28 August. According to the company, this is the first time the family includes native vision and the ability to operate a computer. The Chinese outlet QbitAI tested the model inside WorkBuddy, Tencent's agent tool, and concludes that it delivers complete tasks very well. The account should be read for what it is: a hands-on report by a Chinese specialist outlet, on tasks chosen by its author. It is not an independent measurement.

What Tencent announces with Hy4

A multimodal language model is a model that can process both text and images. According to QbitAI, Hy4 has 770 billion parameters in total, of which 49 billion are activated for each request, which corresponds to a mixture-of-experts architecture. One caution: the site Digital Applied says these figures belong to the preview, and that as of 5 November Tencent had not published a model card for the final version. The model is offered in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, the Yuanbao assistant, the WeCom assistant and the ima app, and its API is available through OpenRouter, Hermes and OpenCode, per QbitAI. The price announced by Tencent, as reported by Digital Applied, is $0.834 per million input tokens, $2.501 per million output tokens and $0.042 for cached tokens, the same as the preview. QbitAI adds that WorkBuddy, in its Chinese and international versions, offers a free trial until 15 November at 11:59 pm.

• Finance evaluation run with Tencent Financial Technology: 213 professionals, 579 tasks; Hy4 scores 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3 (Tencent's internal figures).
• Speculative-type errors cut from 23.1% to 5.6% on Tencent's evaluation sets.
• Hallucination rate: from 15.76% to 5.45% in Tencent's announcement as reported by Digital Applied, but from 12.95% to 7.2% in QbitAI's article. The two sets of figures do not match.
• Vision: locating a roof element of about 2 metres in an aerial image of 2.8 billion pixels, according to QbitAI.
• Research: internal TenureTrail evaluation, 123 executable tasks across 7 disciplines.

What the QbitAI test actually shows

The first challenge given to Hy4 in WorkBuddy was an interactive 3D mechanical keyboard in the 75% layout, with 84 keys and six layers (keycaps, switches, plate, circuit board, battery, case), which can be taken apart with the mouse and animates key presses. The second was to rebuild a retro music player from a single screenshot. The reporter rates visual fidelity at nearly 99.9%, a subjective judgement, and spots only one difference: of the two navigation buttons, one is round and the other square. The model also generated three test tracks using the browser's Web Audio API. According to the execution log cited by QbitAI, it launched a headless Chrome browser to click every button: 48 of 51 tests passed on the first run, one edge case was fixed (moving to the next track after the last one), then 59 tests passed with no console errors.

Hy4 then built a 3D amusement park from a floor plan, with an animated carousel and Ferris wheel, a minimap and an automatic tour. It produced a 16-page presentation on a portable projector (15 were requested), drew the diagrams as vector code and kept the charts editable. Noticing that a block of text overflowed on page 4, it redid the layout before moving on. For a dark template, it tested a single page before applying it to the rest. Finally, it delivered a batch image-processing tool called ATELIER, with parallel processing, before-and-after comparison and ZIP export, after studying open tools such as Squoosh.

The last demonstration was computer use: Hy4 opens JS Paint in Chrome, picks tools and colours, then copies a reference image with the mouse, all recorded. What QbitAI retains is mainly a method: work in small trials, verify, then extend. These demonstrations were chosen by the tester. The article provides no reproducible protocol, no cost, no task duration and no failure rate over a large number of attempts.

Open source or not yet: what remains to be checked

QbitAI's headline places Hy4 among the most powerful open-source models. Yet according to Digital Applied, the weights of the final version were not public on 5 November: the Hugging Face page linked from OpenRouter returned a 401 error and the GitHub repository a 404. Tencent writes that it plans to open the weights and publish a system card at that point; QbitAI speaks of an upcoming release. The only downloadable Hy4 model is the August preview, under the Apache 2.0 licence according to the Hugging Face card cited by Digital Applied. The licence and the parameter count of the final version have not been published. For now, the phrase therefore describes a promise, not a file you can download.

The evaluations cited are all internal to Tencent or supplied by partners. The testimonials from Manus (nearly 97% success on document tasks), RabyteTech, WorkBuddy and Yuanbao are published by Tencent, and two of the four come from its own products, Digital Applied points out, adding that it found no independent reproduction. QbitAI also reports that Hy4, with a Peking University team, found a counterexample to a mathematical conjecture and checked it in Lean: we could not verify this. A lighter version, Hy4 flash, is announced without a date.

What this changes for companies

For a company, Hy4 is first of all a candidate to test, not a choice to lock in. The announced price makes a trial cheap, and the WorkBuddy free trial runs until 15 November. But until the weights, licence and system card are published, planning to host it on your own servers would be premature, and API access means checking where the data travels. The safest course is to measure the model on your own office and development tasks before believing vendors' rankings.