Hy4 flash
AvailableTencent's low-cost Hy4-series model for everyday productivity (research, slides, spreadsheets), launched Nov 16 2026. MoE with about 220B total parameters per the model card (155.5B backbone plus 63.5B Engram n-gram embeddings, ViT and drafter; Tencent's press release says 224B) and about 12B active (11.3B text + 0.43B vision), hybrid sparse/sliding-window attention, LatentMoE (16 of 512 experts) and a 1M-token context; text and image input. Available in WorkBuddy, Yuanbao and Tencent Cloud TokenHub at $0.07 / $0.281 per 1M input/output tokens ($0.003 cached input), plus OpenRouter. Open weights followed on Nov 17 on Hugging Face, ModelScope, GitCode and GitHub (Hy4-flash, Hy4-flash-FP8 and the pre-trained Hy4-flash-base, ungated) under the Tencent Hy License 1.0, an Apache-2.0-derived license with extra terms for large Model-as-a-Service businesses. Vendor-reported figures.
Specifications
- License
- Open weights · Tencent Hy License 1.0
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 220B · 12B active
- Context window
- 1M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- $0.07 / $0.281
- Modalities
- TextVision
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.