MiniMax-H3-Qwen3-VL-NVFP4

MiniMax-H3-Qwen3-VL-NVFP4

An NVFP4 (4-bit) Qwen3-VL-32B text encoder for MiniMax-H3, made with convert-to-quant. It is a drop-in replacement for the H3 conditioning text encoder (ComfyUI CLIPLoader, type minimax). The H3 video weights are unchanged — only the text encoder is quantized to 4-bit, to bring the full H3 stack onto consumer GPUs.

Quantization

• Quantized with ctq --nvfp4 --qwen_vlm --comfy_quant (weight-only PTQ, comfy_quant layout).
• Protected (kept bf16): token embeddings, first & last language layers, and the full vision tower (--qwen_vlm filter). The middle language-layer linears go to NVFP4 (FP4 E2M1, block 16).
• Size: 16.64 GiB / 17.86 GB (bf16 source is 47.97 GiB) — fits a 24 GB consumer GPU with ComfyUI offload.

What's different

This encoder deliberately keeps more of the network in bf16 than a minimum-size 4-bit encoder does. The --qwen_vlm filter protects — at bf16 — the token embeddings, the first and last language layers, and the entire vision tower, quantizing only the middle language-layer linears to NVFP4. The result is 16.64 GiB, about 2 GiB above the most aggressively quantized community NVFP4 encoders (≈14.6 GiB, e.g. the AWQ-quantized encoder in Comfy-Org/MiniMax-H3), in exchange for leaving the precision-sensitive layers (embeddings, boundary layers, vision) untouched. It is a fidelity-first tradeoff: spend ~2 GiB more VRAM to keep the layers that most affect conditioning at full precision. Everything still fits a 24 GB card with offload.

Quantization fidelity (4-bit vs bf16)

Everything is held constant except encoder precision — the bf16 encoder vs this NVFP4 encoder, same prompts, same seeds, same DiT — over a 16-clip paired set.

Verdict — perceptually and temporally equivalent to bf16:

LPIPS ≈ 0.10 with SSIM ≈ 0.89 means the 4-bit encoder's frames are perceptually the same as bf16's, and the near-identical flicker (13.75 vs 14.12) shows temporal behavior is preserved — the 4-bit conditioning adds no visible shimmer or instability. These are the fidelity conclusion.

FVD — internal relative reference only, not a fidelity verdict. i3d FVD over the same 16 paired clips is 1059.5. This is not a published-scale FVD number: FVD estimates a high-dimensional feature covariance and needs large samples (papers typically use 2048+); at n = 16 the covariance is under-determined, which inflates the absolute value and makes it non-comparable to reported FVD. A meaningful reading would require a same-precision noise floor (bf16 vs bf16 at a different seed); we report the raw figure only for transparency and rely on the perceptual and temporal metrics above.

PSNR over the paired clip is 24.85 dB — low only because a tiny conditioning change makes the diffusion sampler diverge pixel-wise while staying perceptually equivalent, which is exactly why LPIPS/SSIM (not PSNR) are the right lens here.

Verified

NVFP4 runs fastest on Blackwell; on Ada (RTX 4090) a cu130 torch build enables the comfy_kitchen FP4 kernels (otherwise emulated and ~2.3× slower).

Usage (ComfyUI)

1. Put qwen3vl_32b_minimax_h3_nvfp4.safetensors in ComfyUI/models/text_encoders/.
2. Download the companion 4-bit weights:
   • DiT: minimax_h3_ref2va_pruned_nvfp4.safetensors → diffusion_models/ (lilcheaty/MiniMax-H3-NVFP4)
   • VAE: minimax_h3_video_vae_fp16 + minimax_h3_audio_vae_fp32 → vae/ (Comfy-Org/MiniMax-H3)
   • Turbo (optional, 4-step): minimax_h3_fl2v_turbo_4step_v0.1.safetensors → loras/ (Gluttony10/MiniMax-H3-INT8-CONVROT)
3. In the H3 workflow set CLIPLoader to this encoder and UNETLoader to the NVFP4 DiT.

Notes

Weight-only 4-bit quantization of the text encoder only; the encoder's behavior is otherwise that of the base Qwen3-VL-32B / MiniMax-H3 release. Follow the MiniMax Community License and local law.

Credits & license

• Base video model: MiniMax-H3 (MiniMax Community License); encoder base Qwen3-VL-32B; bf16 encoder packaging: Comfy-Org/MiniMax-H3.
• Quantizer: convert-to-quant (silveroxides). Companion 4-bit DiT: lilcheaty; turbo LoRA: Gluttony10.
• License inherited from the base model.