Open Source

Tencent's Hy3 and NVIDIA's audio MoE get precise GGUF quants

New imatrix quants for two massive MoE models with full reproducibility.

Deep Dive

LordNeel has released comprehensive GGUF quant sets for two cutting-edge MoE models. For Tencent's Hy3 (295B parameters, 21B active, 262K context), quants range from Q6_K (226 GiB, effectively lossless) to IQ2_M (90 GiB, fits single 96GB GPU). The Q4_K_M at 167 GiB is recommended as a default. All quants include imatrix calibration, KL-divergence and top-token agreement metrics measured against BF16 reference, and llama-bench throughput numbers. The quant sets are fully reproducible from the files in the repo.

Meanwhile, NVIDIA's Nemotron-Labs-Audex-30B-A3B (30B hybrid MoE, ~3B active, 1M context) brings audio understanding and generation capabilities. Two tracks are provided: text-only GGUFs for standard llama.cpp use, and audio quants requiring a sidecar with NV-Whisper encoder, causal speech decoder, and enhancement VAE. Text quants show up to 2x generation speed improvement over BF16 (e.g., Q4_K_M at 345 tok/s vs 177). A notable MXFP4_MOE quant reduces size to 16.7 GiB with experimental quality. The audio pipeline is not yet integrated into a single GGUF runtime. Both model sets are important for democratizing access to large MoE models through efficient quantization, with full benchmarks and transparency.

Key Points
  • Hy3's Q4_K_M quant (167 GiB) fits on 2x 96GB GPUs with 90.0% top-token agreement and 67.3 tok/s.
  • NVIDIA's audio model has separate text-only quants (e.g., Q4_K_M at 22.8 GiB, 345 tok/s) and audio quants needing additional components.
  • All quants use imatrix quantization and include KL-divergence vs BF16 reference for reproducible quality assessment.

Why It Matters

Efficient, quantized MoE models enable powerful AI on consumer hardware, bridging accessibility and performance.

📬 Get the top 10 AI stories daily