Hy4 Released: Vision Is Here, but Open Weights Will Have to Wait

Tencent Hunyuan has released Hy4, adding vision to the model. The official tweet positions Hy4 among leading open-source models, citing deployment scenarios including finance, frontend, gaming, office work, and research, while also mentioning security considerations around cyber/biosecurity.

On parameters, Reachable Games' Japanese tea garden test page labels Hy4 as 770B total parameters and 49B active, with a cost of $0.77 per project. The page has different models use the Beast development harness to generate an interactive browser scene, with requirements for threejs, an orbit camera, and mouse control, including stone lanterns, raked sand, granite steps, mossy shrubs, and cherry trees in the scene. Mooncast Productions released the Hy4 version, saying Hy4 Preview's output felt more organic, while Hy4 is cleaner, executes well, and benchmarks also look good.

Another test came from X user Xiaodong and Ai, who used the workbuddy harness to have the model generate a pelican riding a bicycle. The video shows Hy4 handling objects, motion, and scene composition.

Pricing and open-source specifics are even more worth watching. X user tom_knock says their APEX-Agents test scored 75.2, higher than DeepSeek V4.1 Flash-1017's 72.1/72.9, with mixed results on other benchmarks, and input at about $0.834 per million tokens. Another Korean-language reply posted Tencent Cloud pricing: $0.834 input, $2.501 output, and $0.042 cache per million tokens, with inference fees at 50% off until December 31. The reply also noted that Hy4 weights have not been released yet; currently downloadable is Hy4 Preview released in August, with Hy4 weights and system card planned for later release, and a cheaper Hy4 Flash also on the way.

If these prices and timeline are accurate, Hy4's selling point is bringing vision into the open-source narrative at low cost. But the word "open source" will only count once the weights actually land. Community reaction has also been direct: some ask whether it can replace DeepSeek V4.1 Flash, some ask whether exam cheating would be reported, and others threaten to remove safeguards. Hy4's vision capability has arrived first; safety controversy and the pace of weight releases are still to come.

Official demo entry: hy.tencent.ai