Hy4 Base vs Inkling 2 Mini
Compare documented capabilities and access. A dash means a comparable figure is unavailable.
| Specification | ||
|---|---|---|
| Task | Multimodal pretraining | Multimodal reasoning and coding |
| Published capabilities | Pretrained checkpoint for continued training and adaptation. Tencent’s flagship Hy4 uses a 78-layer language backbone with a separate vision encoder. | Reuses its middle layers three times per token. Thinking Machines reports 48.2% on DeepSWE v1.1. |
| Model size | 770B total / 49B active per token in the language backbone | 26.4B total / 3.4B active parameters |
| Access | Downloadable under the Tencent Hy License | Downloadable weights and Tinker inference and customization |
| Operating limits | — | Native mixed-precision checkpoint occupies approximately 20 GB. |
| Sources | Model weights | Model weights · Additional documentation |
| Inputs | Text, Images, Video | Text, Images |
| Weights | Open weights · Tencent Hy License | Open weights · Apache-2.0 |
| Released | Nov 20, 2026 | Nov 20, 2026 |