
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.
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 |
Hy4 official introduces visual capability for the first time — not simple OCR or image description, but visual reasoning combining perception, reasoning, and tool invocation:
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.
This is the most groundbreaking feature. Hy4 supports computer operation for the first time, flexibly combining CLI and GUI:
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."
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:
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.
Hy4 official is live on Tencent Cloud ADP. ADP provides three layers of support:
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.
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 |
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 |
Tencent handled safety responsibly:
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.
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。