Your privacy choices

Allow optional cookies for referral attribution, visit analytics, and Google Ads purchase measurement.

Back to blog

Tencent Hunyuan Hy4 for Coding and Agents: Official Claims and Third-Party Impressions

Tencent HunyuanHy4Hy4 flashcoding modelagentcomputer use

This article compiles public information. It is not a test run by Guishu Intelligence. Official claims and third-party personal demos are kept apart because their evidence differs in strength. The latter are a handful of personal posts, not systematic tests.

Official claims by focus area

The following comes from Tencent's research page and official WeChat article, and reflects the vendor's position:

  • Computer use: the Hy4 release version supports computer use (CUA) for the first time, covering the command line and graphical interfaces.
  • Targeted training: it was trained for Blender, CAD and long cross-application tasks.
  • Games: Tencent says it deeply supports mainstream engines such as UE, Unity and Godot, and that GameDev and GameLogic improve over the preview by 19.1% and 9.8%. That is a gain over the preview, not an absolute level and not a comparison with other models.
  • Long-horizon and multimodal work: the Yuanbao team says multimodal agent completion is up 36.8% over the earlier VLM (see the WeChat article). This is a product-internal evaluation.
  • Where flash fits: Tencent suggests Hy4 flash for everyday simple front-end work and computer use, and Hy4 for long-horizon software engineering (see the flash research page). The WorkBuddy evaluation says flash's long-horizon coding drops slightly. For more, see Hy4 vs Hy4 flash: How to Choose.

What the independent evaluation says

Artificial Analysis's Intelligence Index v5.0.1 includes coding and terminal-oriented tasks such as Terminal-Bench 4.0 and FrontierCode-AA 2.0. Hy4 scores 35 against a comparison-group median of 43, and the page notes it is slow with very long outputs. A 35 only says Hy4 does not stand out on this set of tasks; it does not mean Hy4 is weak in every scenario. The AA page does not give a separate coding sub-score that this article can cite, so it cannot be used to judge coding strength. See the Hy4 review.

Third-party personal demos (X, 2026-11-17, very small samples)

These are personal posts from the day before flash's weights were released. Each covers one or a few cases and none is a systematic test.

  • @DDJCXX: made a 3D Jiangnan water-town scene with the same prompt and says Qwen4-Max (xhigh) looks clearly better than Hy4 flash (high), giving timings. Original post.
  • @shabnam_774: built a 3D Snake game from scratch with flash and compared it against DeepSeek-V4.1-Flash-1017 using the same prompt. Sample size 1. Original post.
  • @Histonemax: ran 6 cases overnight, is used to working with Opus, and finds flash usable. Original post.

These posts have no shared prompts, scoring criteria or repeated runs, and there is no way to tell whether the authors looked at multiple outputs. The first is a comparison that leans negative on flash, and the other two lean positive or neutral. Taken together they show only that early reactions are mixed.

If you test it yourself

  1. Use a small task you can verify, such as "the door opens only after you pick up the key, and state resets on restart".
  2. Run two models with the same prompt and the same reasoning level, and record the time.
  3. When checking computer use, limit permissions and scope, and review the log of each step.
  4. Run it at least 3 times before judging, because single generations vary widely.

For how to use the model inside products, see the WorkBuddy and CodeBuddy guide. For more reactions, see User Feedback and Limits, and for flash specs see the flash review.

Guishu Intelligence has not tested the tasks described here. An upstream release does not mean the model is available on this site.

Main references

Checked on 2026-11-18.