Your privacy choices

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

Back to blog

Tencent Hunyuan Hy4 Review: Official Claims vs. Artificial Analysis's Score of 35

Tencent HunyuanHy4model reviewArtificial Analysiscomputer use

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

The short version: Tencent positions Hy4 as a flagship built for real productivity. Artificial Analysis (AA), an independent evaluator, gives it an overall score below the median of its comparison group, and notes it is slow and verbose. These two kinds of source differ and should not be merged into one number.

Basics

  • Vendor: Tencent Hunyuan (Tencent Hy / Tencent Hunyuan). Display name Hy4, API ID hy4.
  • Release: the Tencent research page and the Tencent press release give 2026-11-04. One entry in the research list shows 11-03, so the listed date may be off by a day.
  • Specs: TokenHub's Hy API Guide lists a 1M context window, 960k maximum input, 64k maximum output, text and image input, text output, and reasoning support.
  • Parameter count: the research page does not state total or active parameters for the release version. 770B / 49B are figures for the earlier Hy4 preview and must not be applied to the release. Some social posts attach them to the release; that is not a first-party number.

Official claims

Everything in this section comes from the Tencent research page, the press release and the official WeChat article. These are vendor statements.

  • New capabilities: compared with the preview, the release adds vision understanding for the first time and supports computer use (command line and GUI) for the first time. Tencent also says instruction following, long-document understanding, long-horizon tasks and deep research improved. It was trained specifically for Blender, CAD and long cross-application tasks.
  • Finance crowd test: organized by Tencent FiT, with 213 experts and 579 questions. Official scores: Hy4 3.04, DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68, GLM-5.3 2.68. Tencent calls Hy4 first among open-source models. The test was led by Tencent and is not an independent leaderboard.
  • Partner internal evaluations: WorkBuddy's overall evaluation gives an average score of 83.18 and a pass rate of 96.15%, and Tencent says Hy4 ranks first among Chinese models in 9 of 12 scenarios. Manus says office-document task completion is close to 97%. Xuntu Technology's IRAB score is 61.63; the official site says this is 12.03 points above the preview (the body text elsewhere also says "12.03 percentage points", and this article follows the site's wording).
  • In-house search test: hallucination rate falls from 15.76% to 5.45%, and guess-error rate from 23.1% to 5.6%.
  • Yuanbao internal evaluation: 65.25 rises to 72.53.
  • Games: GameDev and GameLogic improve by 19.1% and 9.8% over the preview.

Tencent also published several public benchmark charts. This article does not rely on the specific scores in those charts. To check the originals, open the Tencent Hunyuan research page.

Independent evaluation: Artificial Analysis score of 35

Data from the AA page for Hy4, checked 2026-11-18, under Intelligence Index v5.0.1:

Item Hy4 Comparison-group median
Intelligence Index v5.0.1 35 43
Output tokens used in evaluation 160M 65.1M
Time per task 13.5 minutes 11.5 minutes

The same page lists time to first token of 117.82 seconds and output speed of 96 tok/s, labels the model Proprietary, and describes it as very verbose. The index is built from tasks including AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt and Terminal-Bench-Science.

Chinese media reports relaying AA say Hy4 ranks second among models from Chinese labs, behind Kimi K3.1 Pro (46). That is secondhand and does not mean a global ranking; the figures on the AA page take precedence. The Decoder's report (secondhand) gives the AA median as 30 and says Hy4 ties with Grok 4.7. That conflicts with the 43 on the AA page. This article could not confirm the tie on the AA page and does not use that median.

The gap between "strongest open-source" and the weights

Tencent's press release of November 4 carries the headline wording "joining the ranks of the world's leading open-source AI models". As of 2026-11-18, Tencent says only that Hy4 weights will be released in due course (see the official open-source post). The weights that have been released are for Hy4 flash, on November 18. AA lists Hy4 as Proprietary, which fits that status. So on the day of the announcement, the "leading open-source" line was a forward-looking statement, and readers should not treat it as a conclusion based on downloadable, reproducible weights. The system card is also planned to come with the open-source release, and this article has not read a full system card.

Tencent also acknowledges on the research page 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.

How to read these numbers

  • Official evaluations are mostly in-house sets or partner scenarios. They show what Tencent wants to emphasize, and are not rankings.
  • AA is an independent source, but it is one composite index. A 35 below the median says Hy4 does not stand out on this set of tasks. It does not mean Hy4 is weak in every scenario.
  • Being slow and verbose has a concrete effect on long agent tasks: time and token use grow together. Short tasks such as everyday documents, retrieval and simple front-end work are less affected. For long-horizon software engineering, test on your own tasks first.
  • To pick a version for everyday use, see Hy4 vs Hy4 flash: How to Choose. For coding and agents, see Tencent Hunyuan Hy4 for Coding and Agents: Official Claims and Third-Party Impressions, and for early user reactions see User Feedback and Limits. For product entry points, see the WorkBuddy and CodeBuddy guide. For the earlier generation, see the Hy4 preview guide; its 770B / 49B applies to the preview only.

Unverified items and limits

  • Total and active parameters of the release version: not published by Tencent.
  • Hy4 weights and system card: not released as of 2026-11-18.
  • This article involved no testing of our own. AA data reflects the page as checked on 2026-11-18 and may be updated later.
  • An upstream release does not mean the model is available on this site.

Main references

Checked on 2026-11-18.