Hy4 flash Base vs DaisyRose-27B
Compare documented capabilities and access. A dash means a comparable figure is unavailable.
| Specification | ||
|---|---|---|
| Task | Multimodal pretraining | Agentic coding |
| Published capabilities | Pretrained checkpoint for continued training and adaptation. The Flash architecture uses a sparse Engram memory module. | Post-trained from Qwen3.8-27B on 1,200 agent tasks. RadixArk reports 12% pass@1 on Terminal-Bench 4.1. |
| Model size | Approximately 220B total / 12B active per token | 27B parameters |
| Access | Downloadable under the Tencent Hy License | Public BF16, FP8 and NVFP4 checkpoints |
| Sources | Model weights | Publisher documentation |
| Inputs | Text, Images | Text, Images |
| Context window | — | 262,144 tokens |
| Weights | Open weights · Tencent Hy License | Open weights · Apache-2.0 |
| Released | Nov 18, 2026 | Nov 18, 2026 |