What Is Tencent Hy4? Vision, Computer Use, $0.83 Price
The short answer
Tencent Hy4 is the finished version of Tencent’s open-weight flagship, released November 4, 2026. It keeps the Hy4 preview’s architecture (a 770B-parameter mixture-of-experts with 49B active, per the preview’s model card) and price, and adds three things the preview lacked: image understanding, computer use (command-line and graphical interfaces) and trained skills for professional software such as Blender and CAD. Weights are promised “in the coming days”.
Key facts (as of November 6, 2026)
| Tencent Hy4 | |
|---|---|
| Released | November 4, 2026 (preview: August 28, 2026) |
| New capabilities | Visual understanding, computer use, GUI + CLI control |
| API price (list) | ≈ $0.834–$0.90 input / $2.50–$2.70 output / $0.045 cached per 1M tokens |
| Launch discount | 50% off on Tencent Cloud TokenHub through December 31, 2026 |
| OpenRouter | tencent/hy4, $0.417 / $1.25 per 1M (discounted), 1,048,576-token context, FP4 |
| Weights | Not yet published; preview weights at tencent/Hy4-preview |
| Smaller sibling | Hy4 Flash, “scheduled for release soon” |
What changed from Hy4 preview
- Vision and computer use. Fine-grained image work, complex charts and visual maths reasoning, and agent control of terminals and desktop apps.
- Fewer hallucinations, by Tencent’s internal tests: hallucination rate down from 12.95% to 7.2%, and errors from unsupported assumptions down from 23.1% to 5.6%.
- Game and front-end work: +19.1% on Tencent’s GameDev and +9.8% on GameLogic benchmarks versus the preview, with Unreal Engine, Unity and Godot integration.
- Finance: in a blind evaluation by 213 financial professionals on 579 tasks, Tencent reports Hy4 3.04 versus DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68 and GLM-5.3 2.68. All of these are vendor-run evaluations; no independent leaderboard numbers were available on November 6.
- Safety framing: offensive exploitation material was excluded from training, and Tencent’s Xuanwu and Keen labs found “no significant concerns” about offensive cyber capability in the areas they tested.
How it compares on price
At list price Hy4 is cheaper than most open-weight frontier models on Together AI: Qwen3.8 Max $2/$6, GLM-5.3 $1.40/$4.40, and Muse Spark Open $1.25/$4.25 per million input/output tokens. Until December 31 the 50% discount puts it at roughly half of those again. Check the full table of current API prices before routing work to it; for the earlier release see what is Tencent Hy4 preview.
Should you use it?
- Try it for agentic workflows that mix screenshots, charts and tool calls on a budget, and for Chinese-language enterprise work, where Tencent reports first place in nine of twelve WorkBuddy scenarios in mainland China.
- Wait if you need to self-host: the weights are not out, and a 770B MoE needs a multi-GPU server even quantised.
- Benchmark it yourself on your own tasks: the release numbers are Tencent’s.
Last verified: November 6, 2026.