Tencent releases Hy4 model with image understanding and computer use via Tencent Cloud and OpenRouter; open weights promised later
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Tencent released the finished Hy4 language model (about two months after the preview). It adds image understanding and computer use (trained on tools like Blender and CAD), and is available through Tencent Cloud, OpenRouter and Tencent apps such as Yuanbao. Pricing is unchanged from the preview (about $0.83 per million input tokens, $2.50 per million output tokens). Weights are not yet released, with no date given, and a detailed safety report is due with them. Tencent also says Hy4 with the Hyra 2.0 agent harness improved internal training throughput by about 20% and trained a version of TypeSafe AI's Jev model. It is releasing two specialised biology and physics models as open models. A faster Hy4 flash variant is planned.
Entities: Tencent, Hy4, Hyra 2.0, Tencent Cloud, OpenRouter, Yuanbao
0 primary
What happened
Tencent has released the finished version of its Hy4 language model, about two months after a preview. It adds image understanding and computer use, and is available through Tencent Cloud, OpenRouter and Tencent apps such as Yuanbao. Pricing is unchanged from the preview, at about $0.83 per million input tokens and $2.50 per million output tokens. The weights are not released, no date has been given, and a detailed safety report is promised only alongside them. Tencent says it will release two specialised biology and physics models as open models, and plans a faster "flash" variant.
Why it matters
Developers and enterprises can use Hy4 now through OpenRouter or Tencent Cloud at stated prices, so it is a usable option for cost-sensitive work, especially where a Chinese-language model is wanted. Independent testing by Artificial Analysis puts it at 35 on its index, second among Chinese models, which makes it a credible but not leading choice. It is well behind the top commercial models on hard maths and coding, so it is unlikely to change decisions for teams already using those. The practical impact is modest and mostly about price and choice, not capability.
What is noise
Calling Hy4 "among the strongest open-weight models" is premature when the weights are not out and have no release date, so nobody outside Tencent can check or run it themselves. The finance results, the drop in web-research hallucination (about 16% to about 5%) and the roughly 20% training-speed gain are Tencent's own internal figures, not independently verified. The line about helping speed up its own AI development sounds bigger than a 20% throughput gain on one internal setup supports.
Watch next
- 01Whether Tencent publishes the Hy4 weights and the promised safety report, and under what licence and date. Without them, the open-weight claim stays unproven.
- 02Independent results for the computer-use and image features, plus updates to Artificial Analysis rankings. Check whether Hy4 stays second among Chinese models against Kimi K3.1 Pro, Deepseek and Zhipu.
- 03Real usage and pricing on OpenRouter, and the release of the faster Hy4 flash variant. Look for whether adoption grows or rivals cut prices in response.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680