Tencent released Hy4 on November 4, the full version of the flagship model its Hunyuan team previewed in August, and for the first time in the line the model can look at images and operate a computer. The 770-billion-parameter system is live in Tencent’s own products and through outside platforms, with its open weights promised but not yet published. Tencent presents the release as its entry into the top tier of open models, while openly conceding that a significant gap to the best closed systems remains.
What is new
Vision is the first change, and it reaches beyond describing a picture. Tencent says Hy4 can read charts and documents, reason across several images, and handle visual mathematics, including fine detail in large images, where small features decide whether a result is useful.
One demonstration cited in coverage of the launch had the model locating a roof fitting about two metres across in a 2.8-billion-pixel aerial photograph, a task that tests patience with scale as much as eyesight, and one that a text-only model could not attempt at all.
Computer use is the larger change, because it turns reading into doing. Hy4 can work through a command line and through on-screen applications, and Tencent says it trained the model on professional software such as Blender and computer-aided design tools.
The pitch is aimed at office work, financial analysis, software development, and automation that spans more than one program, with training on long tasks that cross several applications. It is a workplace proposition, built for jobs that span tools, not a chatbot trick.
The model is available now as a service rather than as a download. It is live in Tencent’s Yuanbao chatbot, WorkBuddy, CodeBuddy, ima, and the WeCom assistant Dayuan, and on Tencent Cloud’s TokenHub, with outside access through OpenRouter and other platforms.
WorkBuddy and CodeBuddy carry a limited free quota from November 4 to November 15, and TokenHub calls are half price until December 31. Those dates are the launch offer, not a permanent price, and they set a window for testing.
What Tencent claims, and what it concedes
Tencent’s own benchmark chart puts Hy4 close to the leaders on one test and well behind on others. On OSWorld, a test of operating a computer, Hy4 scored 78.9, against 81.8 for Anthropic’s Claude Opus 5.5 and 75.2 for OpenAI’s GPT-6.1 Sol.
On Terminal Bench it scored 44.4, against 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol, and on ProgramBench it scored 30.0, against 56.5 for Claude. All of these figures are Tencent’s, and several come from tests it built itself.
The company is frank about the distance that remains. Its launch post includes a section on Hy4’s limits, saying the model still has a significant gap to the top closed models, especially on its hardest internal tests of frontier mathematics, long software projects, model development, and visual questions.
Tencent also says it is laying groundwork for recursive self-improvement, and that Hy4, working with mathematicians at Peking University, found a counterexample to a published conjecture and verified the proof in Lean. That claim, like the scores, is Tencent’s own, and it has not been independently checked.
On safety, Tencent says it kept defensive security material in training, such as proof-of-concept generation and bug fixing, but removed exploit, reverse-engineering, and penetration-testing material to avoid amplifying offensive capability. That filtering is a deliberate limit on offensive capability today.
Its Xuanwu and Keen security labs found no offensive cyber risk, it says, and Hy4 scored 30.4 out of 100 on its biology-lab test, below what Tencent calls the high-risk threshold, a judgment that rests on Tencent’s own test design and scoring. These, too, are company findings, not independent audits.
Why It Matters
Hy4 matters most as a statement about where Tencent competes. The company spent years seen as lagging in the AI race, then rebuilt its model infrastructure in February and has shipped a version every two months since, preview first and release after user feedback.
Hy4 is the version that combines visual perception, reasoning, and tool use in one system, which is the combination enterprise software needs if a model is to do office work rather than merely answer questions about it, and it is why this launch lands differently from a simple scale-up.
The open-weight question is where the release is still incomplete. Tencent says it plans to release Hy4’s weights, with vLLM and SGLang support, alongside a fuller system card, but it has given no date and has not said which licence the full model will use, unlike Mistral’s Large 4 release, which spelled out its licence terms.
Until the weights appear, nobody outside Tencent can download, run, or fine-tune Hy4 itself, and the “open” label describes an intention, not a file anyone can inspect. A lighter Hy4 Flash, focused on speed and cost, is also promised soon, without a date.
Our read
The honest admission is the most interesting part of this launch, because it reframes what “top tier” means. Tencent is claiming a place among the strongest open models, not parity with the best closed ones, and its own chart shows why.
Hy4 is competitive where the task is operating a computer, and clearly behind where the task is hard reasoning or long coding. That split fits the product strategy, since Tencent’s first customers are its own workplace tools, where a model that can see a spreadsheet and click through an application is worth more than one that aces a maths contest.
The benchmark comparisons also need a warning label that Tencent itself supplies. When a vendor reports wins on its own tests, the numbers are a progress report, not a verdict, and the finance test that Tencent ran with 213 experts across 579 real tasks is best read the same way.
In that test Hy4 averaged 3.04 out of 4, against 2.97 for DeepSeek and 2.68 for both Kimi K3.1 Pro and GLM 5.3, but the test was run with Tencent’s technology unit. The gap to Claude and GPT on the hard tests is the number to watch when independent evaluations arrive, and it is the number Tencent has conceded.
For buyers, the practical question is narrower than the headline. Hy4 is usable today through Tencent’s products at the preview’s prices, 6 yuan per million input tokens and 18 yuan per million output tokens, with a TokenHub discount to year-end.
The decision to build on it should therefore rest on how it performs in a specific workflow, not on its place in a ranking, and on whether the promised weights and system card arrive in time to make the “open” claim real.
Background
Hy4 succeeds the Hy4 preview from late August, and its 770 billion parameters are a 2.6-fold increase on the 295 billion of Hy3. Its rivals in China’s open-model race include DeepSeek, Z.ai, Moonshot, and Meituan’s visual LongCat model, whose Kimi K3.1 Pro trades wins with Hy4 on Tencent’s own tests.
Tencent has said its next model, Hy5, will scale further, and the company increased capital spending to fund its AI efforts. Hy4 is the clearest test yet of whether that spending has bought it a seat at the open-model table, and the answer will come from independent tests.
Source: SCMP report on the launch and TechNode launch coverage.
0 Comments