llama.cpp b9993 adds Hunyuan 3 support with MTP speculative decoding
Build b9993 merges PR #25395, adding native hy_v3 architecture support so Tencent's Hunyuan 3 MoE runs locally in llama.cpp, with MTP speculative decoding included.
llama.cpp build b9993, released July 13, merges PR #25395 and lands native support for Tencent’s Hunyuan 3 (architecture name hy_v3). The implementation covers the full MoE decoder stack: per-head Q/K RMSNorm, a sigmoid router with expert selection bias, an always-active ungated shared expert, and leading dense blocks. MTP speculative decoding is included from the start.
GGUF quants are already on Hugging Face. The satgeze/Hy3-1M-GGUF repo has what it describes as the first GGUF conversion of Hy3, a 299B MoE model with roughly 17B parameters active at inference time, extended to a 1M-token context with YaRN. The MTP speculative decoding story has a catch worth knowing: the MTP head is trained single-depth, and without a confidence gate (--spec-draft-p-min 0.75) speculation becomes a net slowdown rather than a speedup. Set that flag or don’t bother with speculative decoding. The base model is Apache-2.0.
The implementation in b9993 is ported from charlie12345’s community fork, adapted to current mainline APIs. Credit where it’s due: the upstream PR lists charlie12345 and Piotr Wilkin as co-authors. If you were already running Hy3 on the community fork, you can switch to mainline now without losing MTP support.