A model gets dramatically better and the price stays the same. That is the unusual pitch behind Tencent's Hunyuan Hy4 (混元 Hy4), unveiled on 4 November 2026.
Tencent's Hunyuan lab, part of the company's large model (大模型) push, has spent 2026 rebuilding its training and reinforcement-learning infrastructure. Hy4 is the payoff: the company says it now ranks among the strongest open-weight models (开放权重模型), a first for the Hunyuan series, and that model weights will be released.
The claim is not modest. For most of 2025 and early 2026, Tencent's models trailed the clear front-runners from DeepSeek, Alibaba, and Zhipu on public benchmarks. Hy4 is the lab's attempt to close that gap — and, just as important, to do it without asking buyers to pay more.
Key takeaways
- Hunyuan Hy4 was released on 4 November 2026 by Tencent's Hunyuan (混元) lab.
- It is the first Hunyuan model Tencent positions among the strongest open-weight releases; weights are planned to be open-sourced.
- Hy4 adds native vision understanding and computer-use (电脑操作) for the first time, plus stronger instruction following and long-context comprehension.
- Pricing is unchanged from the previous generation: roughly ¥6 (≈ US$0.85) per 1M input tokens, ¥18 (≈ US$2.54) output, ¥0.3 (≈ US$0.04) cache-hit input.
- In internal office-automation tests it scored 83.18 average and a 96.15% task pass rate, with 9 of 12 scenarios ranked first domestically.
What changed since Hy3
The previous formal step, Hy3, arrived in July 2026 as a 295-billion-parameter mixture-of-experts model with 21 billion active parameters and a 256K context, released under Apache 2.0. Hy4 builds on that base with far larger data and training scale, and — notably — bakes in vision. It can do fine-grained image understanding, chart analysis, and math-from-visuals reasoning.
It also learns to operate a computer: combine command line and graphical interface to handle work and life tasks, with specialized training on professional software such as Blender and CAD and on cross-app long-horizon tasks. This "computer-use" capability is the same family of skill that Western labs have been racing toward, but Tencent is tying it directly to its own office products rather than shipping it as a standalone research demo.
The product co-design angle
Tencent is not shipping Hy4 as a bare model. It is woven into WorkBuddy (and CodeBuddy), the company's agentic office suite, where it lifts document aesthetics, stability, and usability. In a 12-scenario office benchmark the model averaged 83.18 and passed 96.15% of tasks. For finance, Tencent says it gathered expert feedback to tune fund research, securities analysis, and quantitative strategy behavior.
This is the part that separates Hy4 from a paper release. Tencent's advantage is distribution: it can put a new model in front of hundreds of millions of users across WeChat, QQ, and its enterprise tools almost overnight. A capability that lives inside an office suite people already use is more likely to generate the real-world feedback loops that improve models.
Pricing as a statement
Here is the part worth noticing: Hy4's pricing is identical to the prior generation. At roughly ¥6 per million input tokens and ¥18 per million output, with ¥0.3 cache-hit input (≈ US$0.85 / US$2.54 / US$0.04), Tencent is signaling that capability gains should not automatically mean cost jumps.
The flat price matters because the model is more expensive to serve than its predecessor — bigger training, vision, and computer-use all add inference cost. Holding price steady is a strategic choice to win adoption, not an accident of cheaper hardware.
What readers can do now
- If you evaluate open-weight models, put Hy4 on the shortlist for vision + agentic office work — but wait for the actual weight release and an independent benchmark before committing.
- Compare on total task cost, not headline capability. Tencent held price flat while raising capability; that ratio is the real differentiator.
- Try it inside WorkBuddy to see if the computer-use and finance features fit your workflow.
Honest limitations
Key facts here come from Tencent's official launch page (tencent.com) and China Economic Net's technology channel (scitech.ce.cn), a state-affiliated outlet, both reporting the 4 November 2026 release. The 83.18 average score, 96.15% pass rate, and 9-of-12 domestic-first claim are Tencent's internal benchmark figures, not independently re-run here. The "strongest open-weight" positioning is Tencent's; I have not yet seen third-party benchmark confirmations or the published model card at writing time. Pricing is Tencent's list price in RMB, converted to US dollars at roughly ¥7.1 = US$1; actual API pricing may vary by tier and region. I have not compared Hy4 against DeepSeek V4, Qwen, or Zhipu GLM-5.3 on the same tasks in this column.