| license: other | |
| base_model: FastVideo/FastVideo-FastH3-8-Step-V2 | |
| tags: [video-generation, text-to-video, audio, fastvideo, fp8] | |
| # FastH3 V2, 8-step, FP8 | |
| FastH3 V2 (8 distilled steps, video with synchronized audio) with an FP8 transformer for GPUs without FP4 tensor | |
| cores, such as the RTX 4090 and RTX 3090, and for 16 GB and 12 GB memory tiers. | |
| - **Transformer:** FastH3 V2 (50 blocks). Attention and MLP linears in FP8 E4M3 with one scale per output channel; | |
| activations are quantized per token at run time. Everything else stays BF16. | |
| - **Text encoder:** NVFP4 Qwen3-VL trimmed to the 50 layers H3 reads (dequantized per layer on GPUs without FP4). | |
| - **VAE:** LynnReal lightweight video VAE with Kijai's INT8 weights; H3 audio VAE. | |
| - **Sampling:** `fastvideo_inference.json` holds the 8-step DMD schedule, which FastVideo reads automatically. | |
| Source weights: [FastVideo/FastVideo-FastH3-8-Step-V2](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2). | |
| Smaller, faster experimental variant: [FastVideo/FastVideo-FastH3-Trim-8-Step-FP8](https://huggingface.co/FastVideo/FastVideo-FastH3-Trim-8-Step-FP8). | |