Reddit
llama.cpp Merges Native Hy3 Support (PR #25395) — Tencent's 295B MoE Now Runs Locally via 1-Bit/4-Bit GGUF
On July 14, 2026, llama.cpp merged upstream support for Tencent's hy_v3 (Hunyuan Hy3) architecture, a 295B-parameter / 21B-active MoE, so any recent master build can now load it; community GGUF quants (Q2_K, IQ2_M, Q4_K_M) from AngelSlim and others already ship on Hugging Face, with a 1M-context conversion reported running at ~11 tok/s. The PR also wires up MTP speculative decoding (--spec-type draft-mtp, --spec-draft-p-min 0.75), measured at roughly +40% decode throughput on code. This is the local-inference milestone of the week: a near-frontier 295B open model becomes runnable on a single high-memory rig, days after Hy3's Apache-2.0 July 6 release.
↳ Follow the thread