Hy3
Hy3 is a 295B mixture-of-experts language model with 21B active parameters and a separate 3.8B multi-token prediction layer. It follows the Hy3 Preview release with revised post-training.
Model design and access
Tencent expanded reinforcement learning and refined the training data after feedback from product teams. The card focuses on grounding, multi-turn constraint retention and long tasks. Its recommended eight-GPU serving example uses H20-3e or GPUs with more memory.Hy3 official model card
Practical use
Practical guidance. Include incomplete evidence and conflicting user instructions in your review. Check whether a response names the uncertainty and follows the latest valid constraint. Vendor evaluations describe the release but do not establish its accuracy on your documents.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable, API |
| Model identifiers | hy3 |
| License | Apache 2.0Hy3 official model card |
| Context | 256K tokensHy3 official model card |
| Hosted access | TokenHub hy3TokenHub model catalog, updated 5 November 2026 |
Related models
- Hy4 Preview
Tencent · Hy
Hy4 Preview is a text model with a 770B mixture-of-experts backbone and 49B active parameters per token.
Launch date not confirmedPreviewAPI · Downloadable · Provider product - Hunyuan-A13B Instruct
Tencent · Hunyuan-A13B
An instruction-tuned mixture-of-experts language model with 80B total and 13B active parameters.
Launch date not confirmedDownloadable