Tencent's Hy4 Trains a Rival Model That Beats Its Original
By @sharedot · · 6 pages
- AI Frontier
- Open Weight Models
- Agents
Tencent's Hy4 adds vision and computer use, and its Hyra 2.0 harness independently retrained TypeSafe AI's Jev to outperform the original.
Hy4 ships with vision, computer use — and weights still pending
Tencent has released Hy4, the finished version of the model previewed about two months ago.
For the first time in the Hy family, the model understands images — spotting fine details in photos and complex diagrams — and can operate software on its own, trained on professional tools like Blender and CAD programs and on multi-application workflows.
It was also tuned for coding, finance, and office work, with game-engine support for Unreal, Unity, and Godot. The catch: until the model weights are released at an unspecified later date, Hy4 is available only through Tencent Cloud, OpenRouter, and Tencent apps such as the Yuanbao chatbot.
The surprising part: AI that helps build the next version of itself
The development a builder would talk about is not the benchmarks — it is self-improvement.
The Decoder reports that Hy4, working with Tencent's Hyra 2.0 agent system, optimized Tencent's own training pipelines for smaller models, boosting internal training infrastructure throughput by about 20 percent across six steps.
It then independently trained its own version of Jev, a decision-making model from US startup TypeSafe AI — and that variant beats the base Jev model in six out of eight benchmarks, according to The Decoder.
Tencent is releasing two specialized open models for biology and physics created this same way, and plans for Hy4 to keep accelerating AI development inside the company.
The evidence: strong claims, and one very awkward number
Tencent's finance test had over 200 professionals evaluate Hy4 on nearly 600 real-world tasks, where it ranked ahead of DeepSeek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash.
Interesting Engineering reports a 96.15 percent pass rate on Tencent's WorkBuddy evaluations and says internal evaluations cut the hallucination rate from 12.95 percent to 7.2 percent.
But the independent picture is mixed: The Decoder notes that on Artificial Analysis's Intelligence Index, Hy4 scores 35 points — second among Chinese models, behind Kimi K3.1 Pro at 46.
And on the AA-Omniscience test without web search, The Decoder reports Hy4 answers about half correctly and makes up an answer 89 percent of the time when it doesn't know one.
Stakes: cheap tokens, open weights, and tightening export rules
Pricing stays at preview levels: roughly $0.83 per million input tokens and $2.50 per million output, per The Decoder — though Interesting Engineering cites approximately $0.90 and $2.70, so the exact figure differs between publishers.
Tencent itself admits Hy4 falls well behind top commercial models on hard tasks: on Hy-FrontierMath T5, The Decoder reports it scores 13.0 points versus 97.5 for GPT 6.1 Sol and 91.3 for Claude Opus 5.5.
A faster, cheaper Hy4 Flash is planned next, and a detailed safety report will accompany the weight release.
With China's October draft export control framework covering top-10 models above two trillion parameters, the open-weight release timing matters — though the draft reportedly includes a fast-track approval process for open-source releases.