Available on request
Tencent Hy API:
Hy3 or Hy4 preview?
Tencent’s Hy models, from its Hunyuan team, come in two current flagships: Hy3, a tuned production release, and Hy4 preview, a much larger early version. Both are under Apache 2.0. Tell us which one you need and your expected volume, and we’ll talk through adding it to the Xinference Model API.
Side by side
Hy3 vs Hy4 preview
| Detail | Hy3 | Hy4 preview |
|---|---|---|
| Parameters | 295B total, 21B active | 770B total, 49B active |
| Context | 256K tokens | 1M tokens |
| Status | Production release after Hy3 Preview | Early version, shipped with known issues |
| Reasoning control | reasoning_effort: “high” or “no_think” | Defaults to “high”; “no_think” for direct answers |
| Licence | Apache 2.0 | Apache 2.0 |
Sources: Tencent Hy Team model cards for Hy3 and Hy4 preview, checked 23 October 2026. All figures are the model maker’s. Hy, Hunyuan, Tencent and the Hunyuan logo are trademarks of Tencent, shown only to identify the model. Their use does not imply endorsement.
What “preview” means here
Tencent’s own notes on Hy4 preview
It can over-think
Tencent lists spending longer than necessary reasoning through complex tasks as a known issue, which shows up as extra output tokens and latency.
It can over-verify
Tencent also notes a tendency to over-check its own work. For latency-sensitive workloads, Hy3 or the “no_think” mode is the safer choice.
Tencent Hy API questions
Is Hy part of Tencent Hunyuan?
Yes. Hy3 and Hy4 preview are made by the Tencent Hy Team and published under Tencent’s Hunyuan organisation on ModelScope and GitCode, as well as on Hugging Face.
Can I use Hy3 commercially?
Hy3 and Hy4 preview are both released under Apache 2.0, which permits commercial use, modification and redistribution.
What is Hy3 good at?
Tencent highlights coding, office work, financial modelling, frontend design and game development, plus more reliable tool calls and multi-turn instruction tracking than Hy3 Preview.
Is Hy on the Xinference Model API?
On request. The standard models with public per-token pricing are Kimi K3 and DeepSeek V4.1 Flash. For Hy3 or Hy4 preview, talk to us about your volume.
Last updated