Felix Laurent
Tencent AI says Hy4 is out of preview. It calls Hy4 the first Hy model with native vision and computer use. The company posted the news on Nov 4, and says the price stayed at the preview level while the benchmarks improved.
What Tencent announced
Tencent's post is short: "The price stayed in preview. The benchmarks didn't." The Hunyuan team, which builds the models, called Hy4 "firmly among the leading open models". Tencent also says open weights are coming soon, with Hy4 Flash, a faster and cheaper version, right behind.
Tencent previewed the launch at its Global Digital Ecosystem Summit on Oct 29. The preview itself came out in August as open weights. What changed since then is mostly what the model can see and do.
Vision and computer use
The preview was text only, so vision is new in this release. According to Biggo, Tencent says Hy4 reads fine image details, complex charts and visual math problems. For computer use, it can combine command-line and graphical interfaces, and Tencent trained it on tasks in Blender, CAD and across several applications, in a sandbox with hundreds of programs.
Tencent also says it scaled up both pre-training and post-training. It reports gains in instruction following, long context, long tasks and deep research. The figures it gives are relative to the preview.
What Tencent reports
Against the preview, Tencent reports scores 19.1% higher on a game development test and 9.8% higher on game logic. Unite.ai says search hallucinations in Tencent's internal tests fell from 15.76% to 5.45%, and speculative errors from 23.1% to 5.6%.
In a finance evaluation with 213 professionals and 579 tasks, Hy4 scored 3.04 and ranked first among open-weight models. DeepSeek-V4.1-Flash-1017 scored 2.97, and Kimi K3.1 Pro and GLM-5.3 2.68 each. These are Tencent's own evaluations. Tencent's post gives no figures on standard public benchmarks, and Artificial Analysis has no Hy4 listing yet. The price is easier to check.
The price
Unite.ai reports that the listed API pricing is unchanged from the preview: $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens. Biggo also says the official version keeps the preview prices, and adds that Tencent Cloud cuts standard API calls by 50% until Dec 31.
Hy4 is available on Tencent Cloud and OpenRouter, where it lists a context of about 1M tokens. It also runs inside Tencent's own apps, including the Yuanbao assistant, CodeBuddy, WorkBuddy and ima. The weights, though, are not out yet.
What is still open
Unite.ai reports that the model weights and a system card are planned for a later release, with support for vLLM and SGLang. Tencent has not given a date or a license for the Hy4 weights in the sources we read. The preview weights were released under Apache 2.0: 770B parameters in total with 49B active per token. Tencent has not said whether the full model keeps that size. Until the weights ship, the open-weights line is a promise.
Unite.ai also says Tencent acknowledges a gap remains to top proprietary models, especially in research-level math, long coding tasks and advanced visual questions.
So is Hy4 a leading open model? On Tencent's own tests, it beats DeepSeek, Kimi and GLM on finance tasks and cuts its error rates sharply against the preview. It adds vision and computer use at the same price. But the public benchmarks are missing, Tencent admits it trails the best closed models, and the weights that would make it an open model are not out. Until they are, Hy4 is a cheap new Tencent API with an open-weights promise.
