Skip to content
Future Frontpage FF markFFFUTUREFRONTPAGE

The AI developments that belong on page one.

Foundation models

Tencent Hy4 adds vision and computer control, trails leaders on math

The new multimodal model targets coding and finance tasks, but independent tests show gaps in hard math and complex image analysis compared to top rivals.

Key points

  • Hy4 ranks second among Chinese AI models on the Artificial Analysis Intelligence Index with 35 points.
  • Internal tests show hallucination rates in web research dropped from nearly 16 percent to about 5 percent.
  • Model weights for Hy4 are not yet public and will be released at a later, unspecified date.
Tencent Hy4 adds vision and computer control, trails leaders on math
Depth
16 percentInternal tests show hallucination rates in web research dropped from nearly 16 percent to about 5 percent · reported
5 percentInternal tests show hallucination rates in web research dropped from nearly 16 percent to about 5 percent · reported
1 outlet · newest source 21h ago

Tencent released its multimodal language model Hy4, which introduces image understanding and autonomous software operation capabilities. According to The Decoder, the model scores 35 points on the Artificial Analysis Intelligence Index, placing it behind Kimi K3.1 Pro but second among Chinese models.Hy4 is a new AI model from Tencent that can look at images and control software programs by itself.Hy4 introduces multimodal vision and agentic computer control capabilities, though it lags global leaders on complex reasoning benchmarks. The model weights will follow at a later date.

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

The figures

The Decoder reports Hy4 scored 35 points on the Artificial Analysis Intelligence Index. In internal tests, hallucination rates for web research tasks fell from nearly 16 percent to approximately 5 percent. Tencent cited over 200 finance professionals evaluating nearly 600 real-world tasks. The model was tested against rivals including Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash.

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Behind the figures

The Decoder notes that independent tests show Hy4 trails top commercial models like Claude Opus 5.5 on hard math and coding tasks. Tencent stated the model was trained on significantly more data than its predecessor. The finance evaluation involved professionals assessing performance on specific real-world tasks, according to the published report claims according to the published report

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

What it means for your stack

Hy4 is currently available via Tencent Cloud, OpenRouter, and Tencent apps like Yuanbao. The model supports professional tools including Blender and CAD programs. For game development, Tencent adapted Hy4 for Unreal, Unity, and Godot engines. It also handles configuration tasks for the mobile game Honor of Kings, according to the developer team.

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Building on this

Developers can integrate Hy4 for tasks requiring sequential application control, such as operating software on a computer. The model supports coding workflows for website and game building. Finance and office applications benefit from reduced hallucination rates in web research. Until weights are released, developers must use API access through Tencent Cloud or OpenRouter to build on these capabilities.

Evidence: reportedTencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

aifleet-decoder

Why it matters

  • Inference engineers can test autonomous software control workflows using Hy4’s documented support for professional tools like Blender and CAD programs.
  • The drop in hallucination rates for web research offers a quantifiable metric for evaluating reliability in finance and office task automation.
  • The lack of immediate weight release means developers must rely on cloud APIs, affecting deployment latency and cost calculations for production systems.
  • Ranking second among Chinese models but trailing global leaders on math highlights specific capability gaps for complex reasoning workloads.

Topics: models, Tencent, multimodal AI, computer use, inference engineering

Updated

Then / now / next

Then

Tencent’s previous models lacked native image understanding and autonomous computer operation capabilities.

Now

Hy4 introduces vision and computer use, ranking second among Chinese models on the Artificial Analysis Intelligence Index.

Next

Developers should watch for the unspecified future release of model weights to enable local deployment.

Story timeline

  1. Then Tencent’s previous models lacked native image understanding and autonomous computer operation capabilities.
  2. 05 Nov 2026 Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models
  3. Now Hy4 introduces vision and computer use, ranking second among Chinese models on the Artificial Analysis Intelligence Index.
  4. Next Developers should watch for the unspecified future release of model weights to enable local deployment.

For readers

Developers

  • Hy4 supports integration with game engines like Unreal, Unity, and Godot for configuration and bug analysis.
  • Autonomous computer operation allows control of professional tools such as Blender and CAD programs.

Business

  • Finance professionals ranked Hy4 ahead of rivals like Deepseek V4.1 Flash in real-world task evaluations.
  • Reduced hallucination rates in web research may improve reliability for office applications.

Researchers

  • Hy4 scores 35 points on the Artificial Analysis Intelligence Index, trailing Kimi K3.1 Pro.
  • Model weights are pending release, limiting independent architectural analysis.

Numbers in context

16 percentreported5 percentreportedCircle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

When will Hy4 model weights be available?

Tencent has not specified a release date for the model weights, stating only that they will follow at a later date.

How does Hy4 perform on hard math tasks?

Independent tests reported by The Decoder show Hy4 falls well behind top models like Claude Opus 5.5 on hard math and coding tasks.

Can Hy4 operate software autonomously?

Yes, Hy4 is designed to operate software on its own, including controlling professional tools and using multiple applications in sequence.

Connected stories

This storyBaidu's robotaxi data fuels new world model, highlighting physical AI's value and challengesBaidu's robotaxi data fuels new Ricursive releases open-weight Ripple-65B model for chip design automationRicursive releases open-weight R

Reading receipt

  • Hy4 ranks second among Chinese AI models on the Artificial Analysis Intelligence Index with 35 points.
  • Internal tests show hallucination rates in web research dropped from nearly 16 percent to about 5 percent.
  • Model weights for Hy4 are not yet public and will be released at a later, unspecified date.

Still open: When exactly will Tencent release the model weights for Hy4?

Still open: How does Hy4 perform on hard math tasks compared to other specific open-weight models?

Time spent: 0:00 · Depth: Full

Sources

  1. The Decoder

Related stories