FastVideo/Pika-H3-Flex-NVFP4-QAD

Pika H3 Flex NVFP4 QAD

The NVFP4 version of Pika H3 Flex for NVIDIA Blackwell GPUs. It generates video with synchronized audio from a prompt, from first and/or last frames, or from a mix of image, video and audio references. Both Flex adapters are included, and quantization-aware distillation (QAD) against the BF16 Flex model optimizes them for NVFP4 execution.

Pika H3 Flex · FastVideo · Inference guide · Training recipe · FP4-attention variant

Tasks

Task	Inputs	Transformer
T2VA	Prompt	transformer/
I2VA, L2VA, FL2VA	Prompt plus a first frame, a last frame, or both	transformer/
Ref2VA	Prompt plus up to 9 images, 3 videos and 3 audio clips	transformer_ref/

Quickstart

Install FastVideo with FlashAttention-4, then run:

FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN FASTVIDEO_FA4=1 \
python examples/inference/basic/basic_pika_h3_flex.py \
  --model-path FastVideo/Pika-H3-Flex-NVFP4-QAD --nvfp4-dit h3_dit \
  --nvfp4-decoder <snapshot>/vae_decoder_nvfp4/decoder-nvfp4.safetensors \
  --task ref2va --reference image:character.png --reference video:motion.mp4 \
  --prompt "$(cat prompt.txt)" --resolution 768p --duration 5

For keyframe tasks, use --task t2va, i2va, l2va or fl2va, adding --image (first frame) and/or --last-image as the task needs. The inference guide covers every task, canvas sizes and durations, attention backends, and the NVFP4 options.

Contents

Component	Purpose
transformer/	NVFP4 transformer for T2VA, I2VA, L2VA and FL2VA
transformer_ref/	NVFP4 reference-conditioned transformer for Ref2VA
text_encoder/	NVFP4 Qwen3-VL text and vision encoder
vae/, audio_vae/	Video and audio VAEs
vae_decoder_nvfp4/decoder-nvfp4.safetensors	NVFP4 video decoder
scheduler/, audio_scheduler/, processor/, tokenizer/	Pika H3 Flex sampling and input processing
pika_h3_flex_merged.json	Marks that the Pika adapters are already merged into the transformers

The transformers are packed NVFP4 checkpoints and load directly in FastVideo, with no conversion step. Each one already contains its Pika adapter, so FastVideo does not apply it again.

How it was made

• Adapter merge. Each Pika adapter is merged into its MiniMax-H3 transformer at full precision, so quantization starts from the exact Flex weights.
• Calibration. Static activation scales for the feed-forward layers are calibrated on Flex's own sampling trajectories. The clipping is chosen to balance video and audio fidelity.
• Quantization-aware distillation. The NVFP4 student runs the deployed 4-bit weights and activations and is trained against the frozen BF16 Flex model along its six-step sampling trajectory. Only the quantized linear layers are updated.

Compared with post-training quantization alone, QAD brings the model 14-17% closer to the BF16 teacher on held-out inputs, across both adapters. The full recipe, including calibration, training and export, is open source: see the training recipe.

Variants

Model	Attention	Choose it for
Pika-H3-Flex-NVFP4-QAD (this model)	BF16	Output closest to BF16 Flex
Pika-H3-Flex-NVFP4-FP4Attn-QAD	NVFP4 Q/K	Faster long clips

Notes

• NVFP4 kernels require an NVIDIA Blackwell GPU. On other GPUs, use Pika H3 Flex in BF16.
• Quantized output can differ from BF16 output.

License and acknowledgements

This model is a derivative of MiniMax H3 and Pika H3 Flex. It is distributed under the MiniMax H3 Community License Agreement; see NOTICE, THIRD_PARTY_NOTICES.md and licenses/. The Qwen3-VL text encoder is licensed under Apache 2.0.

We thank Pika for open-sourcing Pika H3 Flex, and MiniMax for releasing MiniMax H3.