首页
学习
活动
专区
圈层
工具
发布
社区首页 >专栏 >Hunyuan Hy4 Official: Visual Understanding, Computer Operation, Long-Horizon Tasks — Three New Pieces for Enterprise Agents

Hunyuan Hy4 Official: Visual Understanding, Computer Operation, Long-Horizon Tasks — Three New Pieces for Enterprise Agents

原创
作者头像
克劳德2048
发布于 2026-11-05 08:49:39
发布于 2026-11-05 08:49:39
1440
举报

Tencent Hunyuan Hy4 official version is released today. The official positioning: built for real productivity. Let's unpack what new ammunition the official version gives enterprise agent developers.

Hy4 Official vs Preview at a Glance

Dimension

Hy4 Preview (Aug)

Hy4 Official (Nov)

Improvement

Visual capability

Not supported

First time: image processing, chart analysis, math visual reasoning

0→1

Computer operation

Not supported

First time: CLI + GUI, Blender/CAD specialized

0→1

Long-horizon tasks

Basic

Significantly improved: end-to-end research

Major

Hallucination rate

15.76%

5.45%

-65%

Guess error rate

23.1%

5.6%

-76%

Finance benchmark

Not tested

#1 open-source (3.04)

New

GameDev benchmark

Baseline

+19.1%

19.1%

GameLogic benchmark

Baseline

+9.8%

9.8%

API pricing

¥6 input/¥18 output per M tokens

Unchanged

Same

Three Core Capabilities Decoded

1. Visual Understanding: Agents Can Finally "See"

Hy4 official introduces visual capability for the first time — not simple OCR or image description, but visual reasoning combining perception, reasoning, and tool invocation:

  • Fine-grained image processing: Not just "this is an image," but understanding specific content within
  • Complex chart analysis: Can read Excel screenshots, data visualizations, financial statements
  • Mathematical visual reasoning: Can understand formula images, geometric figures, and reason about them

What it means for enterprise agents: A user sends a screenshot, and the agent understands the content and acts on it. For example, a user screenshots an error message; the agent identifies the error code, queries the knowledge base, and provides a solution — without the user needing to manually describe the problem.

2. Computer Operation: Agents Can Directly Operate Software

This is the most groundbreaking feature. Hy4 supports computer operation for the first time, flexibly combining CLI and GUI:

  • Command-line interface operation
  • Graphical user interface operation
  • Specialized training for Blender, CAD, and other professional software
  • Cross-application long-horizon task support

What it means for enterprise agents: Many enterprise systems don't have open APIs, but agents can operate them via GUI. Agents can enter data into ERP systems, submit workflows in OA systems — no need to modify existing systems. Agents directly "operate like a human."

3. Long-Horizon Tasks & Hallucination: From "Can Run" to "Dare to Deploy"

Metric

Preview

Official

Change

Hallucination rate

从前

5.45%

-65%

Guess error rate

从前

5.6%

-76%

The hallucination drop is a hard prerequisite for enterprise agent production. 15.76% means 1 in 6 answers is fabricated — in customer service, that's 1 complaint per 6 interactions. 5.45% isn't perfect, but it's in the "canary-testable" range.

Long-horizon task improvements demonstrated in research cases:

  • 7M-cell 3D reconstruction: 6.99M cells, 84 unordered 2D slices, 1.5 hours, zero intervention
  • Math conjecture correction: Found counterexample to Méndez–Pinto–Villarreal conjecture, proposed refined conditions, Lean formal verification
  • Training bottleneck identification: Autonomously identified pretraining system bottlenecks, 20.1% throughput improvement

These aren't simple Q&A — the model drives from input to delivery end-to-end. This is the core capability enterprise long-horizon agents need.

ADP × Hy4: Enterprise Agent Landing Solution

Hy4 official is live on Tencent Cloud ADP. ADP provides three layers of support:

Natural Language Building + Agent Harness + Enterprise Service Agent

Natural Language Building: Claw Mode — describe a requirement, get a deployable agent. Now with Hy4's visual capability, agents can process image inputs — users send screenshots to trigger tasks.

Agent Harness: Breakpoint recovery, pause/resume, full-link observability. Hy4's long-horizon task capability can only run stably under ADP Harness — a 20-step task interrupted at step 10 resumes from step 10, not from scratch.

Enterprise Service Agent: Build → Evaluate → Distribute → Govern → Control, with permissions and security policies. Only after Hy4's hallucination drop can enterprises dare to place agents in real production environments.

Hy4 Official on ADP: Typical Scenarios

Scenario

Hy4 Capability

ADP Capability

Combined Value

Customer Service Agent

Long-horizon + low hallucination

Harness breakpoint recovery + WeCom integration

7×24 service, reliable answers

Office Assistant

Human-AI co-writing + visual understanding

Natural language creation + cloud residency

Screenshot triggers tasks, collaborative document editing

Knowledge Q&A

Long-text understanding + deep research

Knowledge base Connector + trace inspection

Multi-step reasoning on complex questions, auditable process

Dev Assistant

Computer operation + code generation

Workflow engine + MCP Connector

Operate IDE + call CI/CD + code review

Internal Operations

Long-horizon + cross-app operation

Canary release + auto-fallback

Scheduled inspection + cross-system operation + exception handling

ADP Double 11 Promotion

  • Hy4 limited-time 50% off: Through December 31, 2026
  • ADP new and existing customer benefits: New customers from 50% off

Where to Access Hy4

Platform

How to Use

Best For

Tencent Cloud ADP

Select Hy4 directly in workbench to build enterprise agents

Enterprise developers, business teams

CodeBuddy / WorkBuddy

Coding and office scenarios

Developers

Yuanbao

Chat experience

Individual users

ima

Knowledge search and reading

Knowledge workers

Tencent Cloud

API access

Teams with in-house dev capability

OpenRouter

API calls

International / multi-model users

Safety & Open Source

Tencent handled safety responsibly:

  • Cybersecurity: Only defensive data included (PoC generation, vulnerability fixing); offensive training content excluded. Audited by Tencent Xuanwu and Keen Lab — Hy4 showed no prominent risk in offensive capabilities
  • Bio safety: Open-ended experiment design capability significantly behind frontier models (Hy-BioExperimentBench: 30.4/100)
  • Open source plan: Model weights to be open-sourced, partnered with vLLM/SGLang community and Inferact Inc. for ecosystem support. Model system card coming soon.

Limitations & Outlook

Tencent frankly acknowledged: while Hy4 is among the strongest open-source models, significant gaps remain compared to top closed-source models, especially on high-difficulty benchmarks in frontier math, long-horizon development, model iteration, and visual QA.

Next: Hy4 flash coming soon, with significantly improved cost-efficiency and inference speed, more frequent product co-design iterations.

Conclusion

Hy4 official's three new pieces — visual understanding, computer operation, long-horizon tasks — each directly addresses real enterprise agent pain points.

Agent couldn't read images? Now it can. Agent couldn't operate software? Now it can. Agent went off track on long runs? Hallucination dropped 65%. Combined with ADP's Harness and Service Agent governance, the bridge from model capability to production capability is built.

50% off through December 31 — now is the lowest-cost window to try Hy4 on ADP.

原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。

如有侵权,请联系 cloudcommunity@tencent.com 删除。

目录
  • Hy4 Official vs Preview at a Glance
  • Three Core Capabilities Decoded
    • 1. Visual Understanding: Agents Can Finally "See"
    • 2. Computer Operation: Agents Can Directly Operate Software
    • 3. Long-Horizon Tasks & Hallucination: From "Can Run" to "Dare to Deploy"
  • ADP × Hy4: Enterprise Agent Landing Solution
    • Natural Language Building + Agent Harness + Enterprise Service Agent
    • Hy4 Official on ADP: Typical Scenarios
    • ADP Double 11 Promotion
  • Where to Access Hy4
  • Safety & Open Source
  • Limitations & Outlook
  • Conclusion
问题归档专栏文章快讯文章归档关键词归档开发者手册归档开发者手册 Section 归档