Your privacy choices

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

Back to blog

Tencent Hunyuan Hy4 User Feedback and Limits: Open-Source Welcome, Personal Demos, Slow and Wordy, and Gaps Tencent Admits

Tencent HunyuanHy4Hy4 flashuser feedbacklimitations

This article compiles public discussion. It is not a test run by Guishu Intelligence, and it is not a survey. It was checked on 2026-11-18, the same day Hy4 flash's weights were released and two weeks after Hy4 launched. Discussion volume is low and much of it is reposts. The categories below summarize opinions and are not benchmarks.

1. Welcome for open weights

The @TencentHunyuan open-source post went up at 04:39 UTC on 2026-11-18. It releases three sets of weights (flash, FP8 and Base) and says Hy4 weights will follow. Many people reposted it. Reposts express welcome for open weights and are not an assessment of quality.

2. Personal demos: highlights and misses

Three posts from November 17, all with very small samples:

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

The only conclusion available is that early reactions are mixed and there is no reproducible comparison. For more, see Tencent Hunyuan Hy4 for Coding and Agents: Official Claims and Third-Party Impressions.

3. 770B wrongly applied to the release version

Some posts present 770B / 49B as Hy4 release specs. Those numbers belong to the earlier Hy4 preview, and the release announcement does not publish total or active parameters. Such claims are unverified folk numbers. See the Hy4 review and the preview guide.

For flash, Tencent (on the research page) says 224B while the open-source model card says 220B, and the two are easy to mix up. See the flash review.

4. Independent evaluation: slow and wordy

The Artificial Analysis page for Hy4 (Intelligence Index v5.0.1): a score of 35 against a comparison-group median of 43, which only says Hy4 does not stand out on this set of tasks and does not mean it is weak in every scenario. Evaluation output was 160M tokens against a median of 65.1M, and the page describes it as very verbose. Time per task was 13.5 minutes against a median of 11.5. This is an evaluation of Hy4, and no flash page was found as of 2026-11-18.

5. Gaps Tencent admits

On the research page, Tencent acknowledges that Hy4 still trails top closed-source models by a wide margin, especially on internal hard-math sets, long-horizon development, model iteration and visual question answering. Tencent also says (in the flash WeChat article) that flash's long-horizon coding drops slightly relative to Hy4.

6. Product-side voices

The improvement figures from Yuanbao (WeChat article), WeCom (Enterprise WeChat), WorkBuddy and WeChat Pay Lab (see the flash WeChat article) are all internal evaluations from within Tencent's ecosystem or from partners, each on its own basis. For choosing between models, see Hy4 vs Hy4 flash: How to Choose. For product usage, see the WorkBuddy and CodeBuddy guide.

7. The system card is not out yet

Tencent says the system card is planned to come with the Hy4 open-source release (see the research page). It was not published alongside the release version, and this article has not read a full system card. On safety, Tencent says offensive cyber data was excluded from training, audits by Xuanwu and Keen say no notable offensive capability risk showed up, and Hy-BioExperimentBench is 30.4/100. These are the vendor's relayed statements without system-card detail, so they cannot be judged independently.

When using this feedback

  • Samples are small and the models are new, so conclusions may change quickly.
  • Personal demos do not represent your tasks.
  • The most reliable approach is to test repeatedly on your own tasks.

An upstream release does not mean the model is available on this site.

Main references

Checked on 2026-11-18.