Hy4 preview

Hy4 preview is the model Tencent's Hy team released and open sourced on August 28, 2026: a 770 billion parameter open-weight model under plain Apache 2.0. It is a mixture-of-experts with 49 billion parameters active per token across 78 layers, 256 routed experts plus one shared expert, top-8 routing, and a separate 10 billion parameter multi-token prediction layer for speculative decoding. The attention design borrows openly from DeepSeek and GLM, using gated DeepSeek Sparse Attention with IndexCache to reuse sparse indices across layers. Context is 1M tokens, reasoning defaults to high effort with a no_think switch, and the vLLM and SGLang recipes ship tool-call parsers. API pricing through Tencent Cloud TokenHub is $0.834 per million input tokens, $2.501 output, and $0.042 for cache hits. The evidence is the weak part. Tencent publishes its benchmark appendix as images rather than a table, and its headline number is an internal blind test in which 163 Tencent experts rated 203 engineering tasks and scored Hy4 preview 2.99 out of 4, against 2.94 for Kimi K3 and 2.92 for GLM-5.3. That is a narrow margin on a panel Tencent picked. Tencent also says the model helped optimize its own training and lifted inference throughput 31.8 percent by tuning operator fusion and communication, which it calls an early recursive self-improvement loop. Treat all of it as vendor-reported, and take the preview label literally: Tencent says more Hy4 models are due soon.