Tencent Hunyuan Hy4-preview runs in vLLM day 0: 770B MoE with 1M context
TencentHunyuan · x · 2026-08-30
vLLM officially announced that Tencent Hunyuan's Hy4-preview is supported from day 0, verified on NVIDIA GPUs.
Architecture highlights:
- 770B total parameters, 49B active per token; 78 layers with the first as dense FFN and the other 77 using 256 routed experts (top-8) plus 1 shared expert
- 1M context, but each query attends to just 2,048 tokens via Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse — only 21 of 78 layers compute their own index
- A native 10B MTP layer (0.7B active, draft depth 3) ships in the checkpoint for speculative decoding
Deployment: the FP8 build runs on 16×B200 or 8×B300; enabling VLLMENABLEHPCOPS=1 activates Tencent's HPC-Ops attention and MoE kernels (in vLLM main since Hy3), combined with MTP speculative decoding and hyv4 tool/reasoning parsers.
More from Infra
- Own your harness, and if possible, own the model layer too — omarsar0 · 2026-08-30
- GLM 5.3 Flash Inference Extremely Slow on Apple Silicon — CentrifugalMalaise · 2026-08-30
- M4 Max Benchmarks: oMLX Wins at Long Context, Prefix Caching is Key — vitordeas · 2026-08-30
- "Sharding Is the Transfer": MindLab Breaks 2M-Context VRAM Wall with 2D KV Resharding — 青稞AI · 2026-08-30
- Oberik: Open-source agent backend solves document parsing and multi-tenancy pain points — kahveciderin · 2026-08-29
- CUDA 'sticky errors' and undefined behavior explained — blelbach · 2026-08-29