cloudyu / hy3-gguf

hy3-gguf — GGUF weights for Tencent Hy3 (tencent/Hy3)

GGUF-format weights for tencent/Hy3 ( HYV3ForCausalLM , model_type: hy_v3 ), a 295B-parameter / 21B-active-parameter Mixture-of-Experts model from Tencent's Hunyuan ("Hy") team ( tencent/Hy3 ).

These files were produced by the hy3 converter ( hy3-convert ) and are meant to be run with the hy3 inference engine , a from-scratch C/Metal/CUDA implementation.

⚠️ This GGUF does NOT work with llama.cpp

Despite the .gguf extension, these files are only usable by the hy3 engine . llama.cpp , ollama , LM Studio , text-generation-webui , koboldcpp , and any other llama.cpp-based tool cannot load these files . Three independent reasons:

1. Unknown architecture. The metadata declares general.architecture = "hy_v3" . llama.cpp only knows hunyuan-moe , hunyuan-dense , hunyuan_vl — loading aborts with unknown model architecture: 'hy_v3' .
2. Custom metadata keys. All hyperparameters use the hy_v3.* prefix ( hy_v3.block_count , hy_v3.expert_count , …), which llama.cpp does not look up.
3. Non-fused expert tensors. Experts are stored one tensor per expert ( blk.N.ffn_gate_exps.0.gate_proj.weight , …1… , … — 46080 tensors), whereas llama.cpp expects experts fused into a single stacked 3D tensor per layer. This is a fundamentally different on-disk layout.

This is a custom GGUF readable only by the hy3 loader. Do not open issues against llama.cpp for these files.

How to run

Use the hy3 engine: https://github.com/yuhai-china/hy3

git clone https://github.com/yuhai-china/hy3
cd hy3
make            # macOS builds the Metal backend automatically

# download a GGUF from this repo, then:
./run_metal.sh -m /path/to/hy3_q4k_mixed.gguf -p "The capital of France is" -experts 8

Testing scope: the hy3 engine's performance work and benchmarks were developed and verified only on macOS / Apple Silicon (Metal backend) , measured on an M2 Ultra (~20–27 tok/s decode depending on -experts ). The CPU and CUDA backends exist in the source but were not exercised as part of that work — treat them as untested.

Files / quantization

The mixed-precision GGUF follows this scheme (see hy3_convert.c ):

Model facts

The engine supports a runtime top-k experts override ( -experts 1..8 ) to trade quality for speed. On a small 13-question code/reasoning eval (greedy, no-think): experts=8 → 10/13 , experts=4 → 7/13 . Default is 8.

Chat template

Hy3 is instruction-tuned and expects the Hunyuan V3 chat format (the hy3 engine applies it automatically; use --raw to bypass). Single user turn, no-think:

<｜hy_begin_of_sentence:opensource｜><｜reasoning_mode:opensource｜>reasoning_effort:no_think<｜hy_User:opensource｜>{prompt}<｜hy_Assistant:opensource｜><think:opensource></think:opensource>

Generation stops on <｜hy_eos:opensource｜> (120025), <｜hy_endofsentence｜> (120001), or <｜hy_EOT｜> (120008).

License & attribution

Weights derive from tencent/Hy3 ; refer to the upstream repository for the governing model license. This is an unofficial community conversion, not affiliated with or endorsed by Tencent.