Tencent's Hy4 adds vision and computer use, with open weights on the way

Hy4 preview arrived on August 28 as a 770B-parameter mixture-of-experts model with 49B active parameters and a context window over 1M tokens[3]. Tencent says the official Hy4 substantially scales up both pre-training and post-training on top of that, and improves instruction following, long-context understanding, long-horizon tasks and deep research[1].

The headline change is that Hy4 can see. It combines image perception with reasoning and tool use to read fine image detail and complex charts, and it can operate computers through both the command line and graphical interfaces[1]. Tencent says it trained specifically on professional software such as Blender and CAD and on long workflows that span several apps[1]. OpenRouter lists Hy4 as taking text and images in and returning text, with up to 65,536 output tokens[2].

• Game development. Working with developers at Tencent Interactive Entertainment Group, Tencent adapted Hy4 for Unreal Engine, Unity and Godot, and reports gains of 19.1% on GameDev and 9.8% on GameLogic over Hy4 preview[1].
• Finance and office work. In a review of 579 real financial tasks by 213 finance professionals, Tencent says Hy4 ranked first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68)[1].
• Fewer made-up answers. On internal tests, the hallucination rate fell from 15.76% on Hy4 preview to 5.45%, and speculative errors from 23.1% to 5.6%[1].
• Research tooling. Paired with Tencent's Hyra 2.0 harness, Hy4 helped develop two models Tencent is open-sourcing: HyCell for single-cell representation and HyPDE for physical field modelling[1].

Tencent says it plans to open-source Hy4's weights and will publish a full system card with its safety evaluations when it does, with vLLM (Inferact Inc.) and SGLang support lined up[1]. It says cybersecurity training data was limited to defensive material, red teams from Tencent Xuanwu Lab and Keen Security Lab found no elevated offensive risk, and the model scored 30.4 out of 100 on its biology experiment-design benchmark, well below its high-risk threshold[1]. IT Home reports a faster, cheaper Hy4 flash is also planned[4].

List API pricing matches Hy4 preview: $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens[1]. At launch OpenRouter was serving it through Tencent Cloud at 50% off, with a 1,048,576-token context window[2].

Hy4 adds another serious Chinese contender to the open-weight frontier alongside DeepSeek, Kimi and GLM, at preview-era prices. Tencent is candid that a notable gap remains with top closed models on research-level maths, long-horizon coding and advanced visual question answering, and its own chart shows Hy4 behind Claude Opus 5.5 and GPT 6.1 Sol on most benchmarks[1]. The real test comes when the weights ship and outside teams can run it themselves.