Your privacy choices

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

Back to blog

Tencent Hunyuan Hy4 flash Review: Open-Sourced on Nov 18, and Why It Is 224B in One Place and 220B in Another

Tencent HunyuanHy4 flashopen-weight modelMoEmultimodal

This article compiles public information. It is not a test run by Guishu Intelligence.

Timeline

  • 2026-11-16: Hy4 flash was announced on the Tencent research page and the official WeChat account (one entry in the research list shows 11-15).
  • 2026-11-18: the official account @TencentHunyuan announced open-sourcing of Hy4 flash, Hy4 flash-FP8 and Hy4 flash-base. The official post also says Hy4 weights will follow, so the Hy4 release version is not yet open-sourced.
  • Distribution channels include GitHub, Hugging Face, ModelScope and GitCode. The license is the Tencent Hy License (see the official post and the model card).

Specs: two sources, two numbers

Source Total parameters Active parameters
Official research page, WeChat article, official X post 224B (including 64B Engram) 12B
Hugging Face model card (same on GitHub) 220B (including 64B Engram) 12B

Tencent does not explain the difference between 224B and 220B, and this article does not reconcile them. When you cite a figure, name its source.

Shared specs come from TokenHub's Hy API Guide: 1M context, 960k maximum input, 64k maximum output, text and image input, text output, reasoning support. The API ID is hy4-flash.

Architecture (open-source model card)

The following comes from the Hugging Face model card and GitHub.

  • The language backbone has 42 layers: 1 dense FFN layer plus 41 MoE layers, with identity Hyper-Connections.
  • The vision encoder is adapted from SigLIP2-So400m-Patch16-Res512.
  • A built-in 5-layer draft model, DFly2 (0.85B).
  • Three structural changes relative to Hy4: hybrid attention (IndexPool interleaved with sliding-window attention at 1:3), LatentMoE, and Engram n-gram embeddings.
  • The model card suggests reasoning_effort of "high" (the default, thinking on) or "no_think", with sampling example temperature=0.9 and top_p=1.0. This is open-source inference guidance, not a parameter recommendation from Guishu Intelligence.

On the research page, Tencent further states that the KV cache is about 1/24 of Hy3's and 1/6 of Hy4's; that post-training self-distills from Hy4, and reinforcement learning cuts end-to-end output length by 50%; and that decoding speed exceeds 150 tps on Tencent's optimized stack. These are official technical claims, and we have not reproduced them.

Official positioning and evaluations

Tencent's division of labor: use flash for most everyday tasks such as deep search, documents, slides, simple front-end work and computer use, and use Hy4, which has a higher intelligence ceiling, for complex work such as long-horizon software engineering. Tencent says flash is slightly better overall than Hy4 preview.

  • WorkBuddy evaluation (WeChat article): everyday-task results are said to be roughly on par with Hy4, long-horizon coding drops slightly, and average task time falls by 70%.
  • Internal blind test: 85 internal WorkBuddy users, 259 tasks across 27 categories. Tencent says user satisfaction is close to Claude Haiku 5.5 and task time is the shortest. This article keeps only that qualitative conclusion. It comes from an official internal blind test and is not an independent ranking.
  • WeChat Pay Lab (WeChat article): in its real-device payment client and static code-scanning scenarios, it says flash matches or exceeds DeepSeek-V4.1-Flash-1017 on average score. This applies only to that lab's scenarios.

Product-side claims

  • Yuanbao: the Yuanbao WeChat account (11-16) says Hy4 flash is now live for all users, with fast answers and deep thinking; users switch modes under the Hy4 model. The official Hunyuan WeChat account relays statements from Yuanbao's product lead: compared with Hy3, agent capability is up 13.6% and human-comparison GSB is up 7.56%; general chat is up 2.52% over GLM-5.3-Flash, most visibly in multi-turn dialogue.
  • WeCom (Enterprise WeChat) (WeChat article): says flash's output speed is 88% higher than Hy4's, and thinking-mode output tokens fall by about 60%.
  • Tencent also says in the WeChat article that flash is live on TokenHub, WorkBuddy, QQ, ima and Marvis, and tencent/hy4-flash has appeared on OpenRouter. For limited-time trials and promotions, refer to the official pages; this article does not show prices or offer figures.

These figures come from product teams' internal measures. Each uses a different baseline and test set, so they cannot be added together or compared with one another.

The independent-evaluation gap

As of 2026-11-18, no Artificial Analysis page for Hy4 flash was found, so there is no independent composite score to cite. The score of 35 on the AA page belongs to the Hy4 release version; see the Hy4 review. Do not read it as flash's result.

Where to go next

To choose between the two, see Hy4 vs Hy4 flash: How to Choose. For early personal demos, see Tencent Hunyuan Hy4 for Coding and Agents: Official Claims and Third-Party Impressions and User Feedback and Limits. For product entry points, see the WorkBuddy and CodeBuddy guide.

Limits

As of 2026-11-18, Tencent has only just announced open weights, and few third-party evaluations that the community can check exist. Every performance figure in this article is the vendor's or a partner's. An upstream release does not mean the model is available on this site.

Main references

Checked on 2026-11-18.