---
license: apache-2.0
base_model: tencent/Hy3
base_model_relation: quantized
pipeline_tag: text-generation
quantized_by: vcruz305
tags:
  - gguf
  - imatrix
  - moe
  - hy3
  - tencent
  - conversational
---

# Hy3-GGUF

imatrix GGUF quantizations of [tencent/Hy3](https://huggingface.co/tencent/Hy3) — **295B total / 21B active** MoE (192 experts, top-8), 80 layers + 1 MTP/NextN layer (3.8B), 256K context.

Quantized day-zero from the BF16 release and **smoke-tested on real hardware before upload** (NVIDIA DGX Spark, GB10). Performance numbers below are measured, not estimated.

## ⚠️ Requires llama.cpp PR #25364 (unmerged)

The `hy_v3` architecture is not yet in llama.cpp master. Until [PR #25364](https://github.com/ggml-org/llama.cpp/pull/25364) merges, build from the PR branch:

```bash
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git fetch origin pull/25364/head:hy3-port && git checkout hy3-port
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j"$(nproc)"
```

These quants were produced at PR head `a4da4b5cfdc4e5fa9def068e216a6e5154f22848`.

## Quants

All quants use an importance matrix (`Hy3.imatrix`, included) computed on a ~63KB diverse coding/reasoning/chat calibration corpus. Files are sharded at ~48GB for HF's 50GB limit — download the whole folder and point llama.cpp at the `-00001-of-*` shard; the rest load automatically.

| Quant | Size | ~BPW | Fits | Notes |
|---|---|---|---|---|
| [Q8_0](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-Q8_0) | 318 GB | 8.6 | server-class | near-lossless |
| [Q5_K_M](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-Q5_K_M) | 212 GB | 5.8 | 2× 128GB-class | |
| [Q4_K_M](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-Q4_K_M) | 181 GB | 4.9 | 2× 128GB-class | recommended dual-node; verified over RPC |
| [IQ4_XS](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-IQ4_XS) | 159 GB | 4.3 | 2× 128GB-class | |
| [Q3_K_M](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-Q3_K_M) | 143 GB | 3.9 | 2× 128GB-class | |
| [IQ3_XXS](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-IQ3_XXS) | 117 GB | 3.2 | single 128GB-class (borderline) | MTP block @ q8_0; best quality-per-GB single-box |
| [Q2_K](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-Q2_K) | 109 GB | 3.0 | single 128GB-class | |
| [IQ2_M](https://huggingface.co/vcruz305/Hy3-GGUF/tree/main/Hy3-IQ2_M) | 100 GB | 2.7 | single 128GB-class | MTP block @ q8_0; smallest tier |

On low-bit IQ tiers the MTP/NextN layer (`blk.80.*`) is kept at q8_0 (`--tensor-type`) — very-low-bit quantization of that block is not possible without imatrix coverage, and this preserves it intact for future speculative decoding support.

## Measured performance (DGX Spark, GB10, 273GB/s unified memory)

Single node, fully GPU-resident, `-fa on`:

| Quant | Gen tok/s | Prompt tok/s |
|---|---|---|
| IQ2_M | 18.5 | 36.4 |
| Q2_K | 18.1 | 35.5 |
| IQ3_XXS | 17.1 | 32.4 |

Dual node (Q4_K_M, 181GB layer-split across 2× GB10 over 200GbE via llama.cpp RPC): **14.0 tok/s gen / 21.9 tok/s prompt**. RPC layer-split adds capacity for bigger quants, not speed — expect single-node-or-slower decode rates.

All smoke-tested tiers (IQ2_M, Q2_K, IQ3_XXS single-node; Q4_K_M dual-node) produced coherent output (code generation + chat), with the Hy3 chat template engaging correctly via `--jinja`. The remaining tiers (Q8_0, Q5_K_M, IQ4_XS, Q3_K_M) were produced by the same verified pipeline but not individually load-tested — report any issues in the community tab.

## Running

```bash
# chat/completion (--jinja is required — the Hy3 template is not natively supported)
./build/bin/llama-completion \
  -m Hy3-Q2_K/Hy3-Q2_K-00001-of-00003.gguf \
  -ngl 99 -fa on -c 8192 --jinja \
  -p "Write a Python function that returns the median of a list." -n 256

# server
./build/bin/llama-server \
  -m Hy3-Q2_K/Hy3-Q2_K-00001-of-00003.gguf \
  -ngl 99 -fa on -c 8192 --jinja --host 0.0.0.0 --port 8080
```

Tips:
- First load of a 100GB+ quant can take several minutes — don't kill it early.
- Constrained on memory? Offload MoE expert tensors to CPU with `--n-cpu-moe N`.
- `--jinja` matters: without it the chat template aborts on current llama.cpp.

## MTP / speculative decoding status

Hy3's MTP (multi-token prediction) tensors are **converted and stored** in these GGUFs, but PR #25364 currently skips them in the forward graph — llama.cpp **cannot yet use them for speculative decoding**. The weights are preserved (q8_0 on low-bit tiers) so existing files become spec-decode-ready if/when the PR adds graph support. No re-download should be needed.

## Provenance

- Source: [tencent/Hy3](https://huggingface.co/tencent/Hy3) (BF16, 597.6GB, 99 shards)
- llama.cpp: [PR #25364](https://github.com/ggml-org/llama.cpp/pull/25364) @ `a4da4b5cfdc4e5fa9def068e216a6e5154f22848`
- imatrix computed on a Q8_0 intermediate (standard practice for models whose BF16 GGUF exceeds node RAM)
- Quantized and validated on a 2-node NVIDIA DGX Spark cluster

Quantized by [vcruz305](https://huggingface.co/vcruz305). Please report issues in the community tab.