Tencent Hunyuan's Hy4-Flash engine PR reveals tensor parallel, MTP drafts, CPU expert lane

TencentHunyuan · x · 2026-11-19

Tencent Hunyuan retweeted a GitHub PR bringing Hy4-Flash to the open-source Project Maya inference repo with a dedicated maya-hy4 engine (NVIDIA, Linux).

Key details: the model has 42 layers plus an MTP block, MLA attention (512-token sliding window, 12 layers with an indexer), 512 routed experts with top-16 routing, 4 hyper-connection streams, and Engram n-gram lookups on 4 layers. The engine supports tensor parallelism (--tp auto|N|0), MTP draft inference, and a CPU expert lane; packing from AngelSlim's GGUF writes geometry/index/tokenizer without copying weights, and the server supports chat, thinking with effort levels, and tool calls.

Original post →

More from Infra

Infra channel →