Tencent Hunyuan Hy4 Official Release: ADP First to Integrate, Enterprise Long-Horizon Agents Ready for
> Today, Tencent Hunyuan officially released Hy4. Compared to the preview open-sourced in August, the official version introduces visual capabilities and computer operation for the first time, significantly strengthens long-horizon tasks and deep research, and cuts hallucination rate from 15.76% to 5.45%. Tencent Cloud ADP has integrated Hy4 on day one — select it directly from the workbench, with a limited-time 50% discount. ## Four Core Breakthroughs The Hy4 preview from August 28 was already impressive — 770B total parameters, 49B active parameters, 1M context, MoE architecture, 2.99/4 in internal blind tests. The official version expands across the board: |Upgrade Direction|Details|Quantified Improvement| | --- | --- | --- | |Visual capability (first time)|Fine-grained image processing, complex chart analysis, mathematical visual reasoning|0→1| |Computer operation (first time)|CLI + GUI combined, specialized training for Blender/CAD and cross-application long-horizon tasks|0→1| |Long-horizon tasks & deep research|End-to-end scientific research, 1.5 hours to complete 3D reconstruction of 7M-cell mouse embryo|Hallucination 15.76%→5.45%| |Finance & office|213 experts, 579 real financial tasks, ranked #1 among open-source|Score 3.04| ### Vision & Computer Tasks: From "Text Only" to "Can See and Operate" Hy4 official introduces visual capability for the first time, combining perception, reasoning, and tool invocation: - Fine-grained image processing and complex chart analysis - Mathematical visual reasoning - **Computer operation**: flexibly combines CLI and GUI, specialized training for Blender, CAD, and other professional software This isn't simple "image description" — the model can understand screenshot contents and operate a computer accordingly. For enterprise agents, this means: agents can operate internal systems, process report images, and execute cross-application tasks. ### Frontend & Games: Functional and Usable Visual understanding significantly boosts Hy4's frontend and game development: - Targeted optimization for complex interactions, business logic, state management, frontend-backend integration - Deep adaptation for UE, Unity, Godot; expanding to multiplayer game development - GameDev benchmark +19.1%, GameLogic benchmark +9.8% ### Finance & Office: Hallucination Plummets, Human-AI Co-Writing Jointly with Tencent FiT, 213 experts (153 from external top financial institutions, 70% with 5+ years experience), 579 real financial tasks: |Model|Score| | --- | --- | |Hy4 Official|3.04| |DeepSeek-V4.1-Flash|2.97| |Kimi K3.1 Pro|2.68| |GLM-5.3|2.68| More critically, reliability: **hallucination rate dropped from 15.76% to 5.45%**, error rate from guessing dropped from 23.1% to 5.6%. For enterprise agents, "not making things up" matters more than "answering smoothly." Hy4 also supports human-AI co-writing — real-time collaborative editing of Word, Excel, and PowerPoint. ### Research: End-to-End Autonomous Execution - TenureTrail benchmark (7 disciplines, 123 executable research tasks) — #1 among open-source - 1.5 hours, zero intervention: 3D reconstruction of 7M-cell mouse embryo - Found counterexample to Méndez–Pinto–Villarreal conjecture, completed Lean formal verification - Autonomous bottleneck identification in model training, 20.1% throughput improvement ## ADP Integration: Three Core Capabilities Tencent Cloud ADP (Agent Development Platform) has integrated Hy4 on day one. ADP provides three layers around Hy4: **Layer 1: Natural Language Agent Building** Describe requirements in natural language; ADP auto-understands intent, matches tools, generates workflows, and produces a publishable agent. Claw Mode enables business teams to build agents with zero code. **Layer 2: Agent Harness — From Demo to Production** This is the key for enterprise agents: breakpoint recovery, pause/resume, full-link observability. What's the biggest fear with long-horizon tasks? Running 15 steps and having to start over when interrupted. ADP's Agent Harness delivers: - **Breakpoint recovery**: Resume from the interruption point without losing completed steps - **Pause/resume**: Pause an agent, inspect intermediate state, then continue - **Full-link observability**: Trace replay, every step's input and output inspectable This layer transforms Hy4's long-horizon tasks from "can run" to "production-ready." **Layer 3: Enterprise Service Agent — 7×24 Always-On** Build → Evaluate → Distribute → Govern → Control, with permissions and security policies. Suitable for: - Customer service - Office assistant - Knowledge Q&A - Dev assistant - Internal operations These scenarios share a common need: 7×24 online, enterprise governance, permission control — exactly where Hy4's long-horizon tasks and ADP's platform capabilities converge. ## Why Hy4 Official Is Significant for ADP On October 29, Dowson Tong previewed Hy4 official at the ecosystem summit. Now it's here. Three changes validate previous analysis: **1. Long-horizon tasks: not just "thinking far" but "finishing the job"** The 7M-cell 3D reconstruction case is proof — 1.5 hours, zero intervention, end-to-end completion. This is exactly what enterprise agents need: not just giving an answer, but autonomously driving from input to delivery. **2. Visual capability opens new enterprise agent scenarios** Agents can read screenshots, process report images, operate GUIs — meaning enterprise system interaction is no longer limited to API calls. For legacy systems without open APIs, agents can operate via GUI, massively expanding agent applicability. **3. Hallucination drop makes enterprises dare to deploy** From 15.76% to 5.45% — the model didn't get smarter, it got more reliable. The biggest fear for enterprise agents isn't "can't answer" but "confidently making things up." This reliability improvement is the critical step toward putting long-horizon agents in production. ## How to Use Hy4 on ADP 1. Log in to ADP workbench (adp.tencent.com) 1. When creating or editing an agent, select Hy4 from the model picker 1. Configure tools and connectors, publish to WeCom or API channels 1. Agent runs cloud-resident, 7×24 always-on Limited-time 50% discount through December 31, 2026. ## Conclusion From August preview to November official, Hy4 spent three months completing the leap from "one of the strongest open-source models" to "enterprise-grade long-horizon agent foundation." Visual understanding, computer operation, long-horizon tasks, hallucination drop — each directly addresses enterprise production pain points. ADP's day-one integration isn't just "one more model option" — Agent Harness breakpoint recovery, full-link observability, and Service Agent governance are what translate Hy4's capability into enterprise productivity. Go to ADP workbench, select Hy4, build a long-horizon agent — that's the best way to test what "long-horizon" really means.
