Tencent releases official version of flagship model "Hy4," its first to understand images; listed API price half that of preview, weights not yet public

Finished version of the preview, first to understand images

China's Tencent released the official version of "Hy4," the flagship of its "Hunyuan" AI models, around midday on Nov. 4 (the afternoon of the same day Japan time). It is the finished version of the "Hy4 preview" released on Aug. 28, with expanded pretraining and post-training. It is the first in the series to support image understanding and also the first with a feature to operate computers by combining command input and on-screen operation. China's IT Home and others reported it.

Besides the API of the company's cloud, "Tencent Cloud," it is now also available on OpenRouter, a U.S. AI model routing service. @gptbotio, an X account that tracks new models, spotted its addition to OpenRouter at around 1:30 p.m. Japan time, and @aitrackerbot reported its release on Tencent Cloud's official API. The weights have not yet been released.

Tencent had announced on Oct. 29 that it would release the official version soon. On Nov. 3, Hunyuan's official X account posted a teaser, but no post from that account announcing the release of the official version could be found as of the afternoon of Nov. 4.

Listed API price half the preview's, Tencent says pricing "unchanged"

On OpenRouter, the price is $0.417 for input and $1.251 for output per million tokens, with Tencent Cloud as the provider. On the same service, the preview is $0.834 for input and $2.501 for output, and the official version is listed at exactly half. Vercel's routing service "AI Gateway" also shows $0.417 for input and $1.2505 for output.

According to IT Home, however, Tencent said of the official version's pricing that it would keep the same API pricing as the preview while greatly improving performance. It did not give specific amounts. Some users have pointed out that the current price is a 50% launch discount and that the regular price is $0.83 for input and $2.50 for output, but this could not be confirmed in Tencent's announcement.

Among Chinese rivals, Moonshot AI's "Kimi K3.1 Pro," released on Oct. 23, costs $2 for input and $10 for output. Meituan's "LongCat-2.5" costs $0.30 for input and $1.20 for output for a limited time, and DeepSeek's "V4.1 Flash" costs $0.30 for input and $1.20 for output during peak hours. Hy4's listed price is cheaper than Kimi K3.1 Pro and more expensive than LongCat-2.5 and V4.1 Flash.

No performance figures, gap with top closed models acknowledged

Tencent said it improved instruction following, long-text understanding, long-running tasks and deep research. It said it also trained the model for specialized software such as the 3D creation tool "Blender" and CAD design software, and for tasks spanning multiple apps. For images, it said the model can be used for processing fine-grained images, analyzing complex charts and solving math problems that include diagrams.

Tencent claimed that for the first time, Hy has joined the ranks of the strongest open models. However, as far as can be confirmed from news reports, no evaluation scores for the official version or the models it was compared with have been given. The company acknowledged that on difficult evaluations it built itself, such as frontier math, long development tasks and questions about images, the gap with top closed models remains large.

Results for the official version also cannot be found on third-party evaluation sites such as Artificial Analysis. For the preview, Tencent had noted in the model card a tendency to think longer than necessary and to check its own answers repeatedly. It is not known whether this was fixed in the official version.

Weights not public, no lightweight version either

The preview's weights were released at launch under the Apache 2.0 license, which permits commercial use. The weights for the official version have not been released, and the only Hy4 models listed on Tencent's Hugging Face page are the August preview and its lightweight versions.

IT Home and others reported that Tencent would release the weights soon, but no timing was given, and this could not be confirmed in Tencent's official announcement. The lightweight "Hy4 flash," announced on Oct. 29, was not released this time. IT Home reported it would be released soon, but this also could not be confirmed.

Offered on its own cloud and apps

The official version is available as an API on "TokenHub," Tencent Cloud's model platform. Tencent Cloud's API can also be called from OpenRouter and Vercel's AI Gateway. The context window is about 1 million tokens, and a single output is up to 64,000 tokens.

Among Tencent's own services, it has been built into the AI assistant "Yuanbao," the coding assistant "CodeBuddy," the work-oriented "WorkBuddy," the knowledge management tool "ima" and "Dayuan," the assistant in the enterprise chat app "WeCom."

Released 6 days after announcement, Alibaba's "Qwen 4" also waiting

Hy4 was released six days after the Oct. 29 announcement. On Nov. 2, steps to run the official version were added to a proposed change to inference software for China's Huawei chips, prompting views that the release was near.

China's Sina Finance reported on Oct. 30 that Alibaba was also expected to release its next flagship, "Qwen 4," the following week. The Qwen 4 name has appeared ahead of time in the settings of a developer tool, but its release could not be confirmed as of the afternoon of Nov. 4.