Tencent has given Hy4 eyes and hands. The company announced the full launch of the model on 4 November, adding image understanding and computer use that the text-only August preview did not have. Weights and system cards, the documents describing how the model was tested for misuse, are promised for “the coming days,” so for now the open-source label rests on a pledge.
On the visual side, Tencent says Hy4 can work on detailed images, read complicated charts, and reason about math problems posed visually. The computer-use feature lets the model operate a terminal as well as point-and-click screens. Tencent says it received specialized training on professional software including Blender and CAD, the tools a 3D artist or an engineer lives in. That choice points at long, multi-step jobs, not one-shot chat answers.
The company’s own numbers lean on its internal products. On WorkBuddy, its workplace assistant, Hy4 scored 83.18 on average with a 96.15 percent pass rate, and Tencent says it took first place in nine of the twelve scenarios it tested in mainland China. A blind test with Tencent Financial Technology used 213 financial professionals (153 of them from outside institutions) and 579 real-world tasks. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Tencent’s own team ran that evaluation, and the margin over DeepSeek is 0.07 points.
The sharper comparison is with Hy4 itself. Tencent reports that the hallucination rate fell from 12.95 percent in the preview to 7.2 percent, and that errors from unsupported assumptions dropped from 23.1 percent to 5.6 percent. Those are internal evaluations. Still, a hallucination rate that falls by nearly half matters most for an agent that clicks through real software, where one invented fact can become a wrong action.
Price did not move. Prices match the preview: about US$0.90 to send a million tokens (the chunks of text a model reads) in, US$2.70 to get a million out, and US$0.045 for a million cached inputs. Developers can reach it through Tencent Cloud TokenHub, and third-party platforms including OpenRouter and OpenCode are carrying the API. A cheaper Hy4 Flash is due soon. Through 15 November, Tencent is giving away WorkBuddy and CodeBuddy, and TokenHub customers get 50 percent off standard pricing until 31 December 2026.
Safety claims come with the same caveat as the performance ones, since Tencent wrote them. The company says Hy4 was trained on defensive security data and that offensive exploitation material was left out. Specialists at two of its in-house security groups, Xuanwu and Keen, reviewed the model and reported no significant concerns about its offensive capabilities in the areas tested. On biology, internal reviewers found strong foundational knowledge but weaker open-ended experimental design than frontier models.
Tencent published these details in a press release on its own site. A model that operates a desktop and has weights anyone can download is a different kind of release from a chatbot, because outside testers can probe it without Tencent’s supervision. If the weights and system cards arrive on schedule, expect independent computer-use tests against the Western agents within weeks, and those results will say more than a figure of 83.18 on Tencent’s own product.
Reported by Tencent in a press release on 4 November 2026.
