Hy4 preview
770B-total, 49B-active open MoE with 1M-token context, Gated DeepSeek Sparse Attention and hyper-connections, under Apache-2.0.
A step up from Hy3 in size, context and data, with 78 layers, 256 routed experts and a 10B MTP layer. Per its model card the architecture is inspired by DeepSeek and GLM. Tencent flags over-long reasoning as a known flaw. Hy4 preview topped OpenRouter token volume in early September (Wikipedia).
- Date
- Friday, 28 August 2026
- Lab
- Tencent
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 770B / 49B | company |
| Context length | 1M | company |
| OpenRouter weekly tokens (week to 2026-09-06) | 14.7T ranked first on OpenRouter, per Wikipedia citing OpenRouter | independent |
Release 2026-08-28 per Wikipedia and OpenRouter listing (HF repo created 2026-08-27 UTC). Benchmark numbers for Hy4 in its card are in images; none recorded. Bloomberg reported Tencent claims it beats Z.ai and Moonshot models (article not opened).
Sources
This record was checked against its sources on 6 October 2026. How we check