Hy3-GGUF

Hy3 — GGUF (imatrix)

High-quality imatrix GGUF quantizations of tencent/Hy3, Tencent's 295B-parameter Mixture-of-Experts model with ~21B active parameters per token.

Produced with a Hy3-enabled llama.cpp branch (build 9909, commit 56142c5f8). Every quant in this repo was measured for quality (KL divergence, perplexity, and top-token agreement against BF16 reference logits) and throughput (llama-bench on 8× RTX PRO 6000 Blackwell) — full raw data is included under benchmarks/.

Model Overview

The GGUFs are converted from the BF16 base model with the MTP/NextN block excluded (--no-mtp) — intentional for normal serving. Q8_0 was intentionally skipped for this release.

Which File Should I Download?

Sizes below are weights only — leave headroom for KV cache and compute buffers.

Download

# Single quant (example: Q4_K_M)
hf download LordNeel/Hy3-GGUF Hy3-Q4_K_M-imatrix.gguf --local-dir ./Hy3-GGUF

Run with llama.cpp

Use --split-mode layer for multi-GPU. (--split-mode tensor crashed in CUDA decode during testing on this Hy3 GGUF path — see Known Limitations.)

llama-server \
  --hf-repo LordNeel/Hy3-GGUF \
  --hf-file Hy3-Q4_K_M-imatrix.gguf \
  -ngl 999 \
  --split-mode layer \
  --tensor-split 1,1,1,1,1,1,1,1 \
  -c 8192 \
  -b 1024 \
  -ub 512

The chat template is embedded in the GGUF, so llama-server / llama-cli chat modes work out of the box.

Quality Metrics

Quality was measured with llama-perplexity --kl-divergence against BF16 no-MTP reference logits generated from tencent/Hy3.

Eval setup

• Corpus: WikiText-2 raw test (llama.cpp scripts/get-wikitext-2.sh)
• Slice: 128 chunks × context 512 (nominally 65,536 tokens)
• Flags: -c 512 --chunks 128 -b 1024 -ub 512 -ngl 999 --split-mode layer --tensor-split 1,1,1,1,1,1,1,1
• The BF16 reference logits file was kept local as a large intermediate and is not uploaded.

Lower KLD, PPL ratio, and RMS Δp are better; higher same top-p is better. "Same top-p" is the percentage of positions where the quant and the BF16 reference agree on the highest-probability token.

Raw quality data: benchmarks/kld_wikitext2_c512_chunks128.json · benchmarks/kld_eval_manifest.json · benchmarks/kld_logs/

Throughput

llama-bench, CUDA backend, 8× NVIDIA RTX PRO 6000 Blackwell Server Edition, -ngl 999 --split-mode layer --tensor-split 1/1/…, prompt 512, generation 128, 3 reps, batch 1024, ubatch 512, F16 KV cache.

BF16 no-MTP GGUF reference size used for compression ratios: 549.66 GiB (590,197,887,712 bytes).

Raw throughput data: benchmarks/summary.json and the per-quant benchmarks/*_llama_bench_p512_n128_r3.json files.

Charts

Production Serving Notes

For quality-first GGUF serving, use Q6_K:

For maximum batched throughput on the tested 8-GPU machine, a separate (non-GGUF) FP8 SGLang profile was faster:

• SGLang 0.5.5 Hy3 backport, TP=8
• FP8 weights + FP8 E4M3 KV cache with calibrated K/V scales
• CUDA graph batch sizes [1,2,4,8,16,32]
• Best measured: 688.76 aggregate output tok/s at concurrency 32 over 128 successful requests

Raw reference: benchmarks/sglang_fp8_tp8_reference.json. Tencent's base model card provides vLLM and SGLang recipes for Hy3.

Quantization Recipe

• Source GGUF: BF16, --no-mtp, --fuse-gate-up-exps (1199 tensors in the no-MTP BF16 GGUF)
• Imatrix: llama-imatrix, 96 chunks, calibration text from hy3_calibration_512.jsonl (uploaded as calibration/hy3_imatrix_calibration.txt)
• Imatrix file: imatrix/Hy3-imatrix.gguf — included so you can reproduce or make additional quants

llama-quantize \
  --imatrix Hy3-imatrix.gguf \
  --leave-output-tensor \
  --token-embedding-type q8_0 \
  Hy3-BF16.gguf \
  Hy3-Q4_K_M-imatrix.gguf \
  Q4_K_M \
  64

The full reproduction runner used for this release: scripts/run_hy3_gguf_stream_upload.sh

Known Limitations

• KLD/PPL numbers come from a bounded WikiText-2 slice (128 × 512 ctx), not a full downstream task harness. MMLU/GPQA/coding evals were not run for these GGUFs; see the upstream task scores on Tencent's base model card instead.
• The GGUFs exclude the MTP/NextN block — intentional for normal serving.
• --split-mode tensor was unstable (CUDA decode crash) on the tested stack for this Hy3 GGUF path. Use --split-mode layer.
• A GPTQ4/SGLang variant was tested separately and classified invalid (generation collapsed into repeated punctuation); those artifacts are not published here.
• Q8_0 was intentionally skipped for this release.

Benchmark Hardware

• 8× NVIDIA RTX PRO 6000 Blackwell Server Edition (~97 GB VRAM each)
• CUDA 12.8, driver 570.195.03
• AMD EPYC 9B14 host, 224 logical CPUs, ~1.5 TiB RAM

Acknowledgements

• Tencent Hunyuan for the Hy3 base model (Apache 2.0)
• The llama.cpp project and the Hy3 support branch these conversions were built on