f1nely/LongCat-Video-Avatar-1.5-okno

LongCat-Video-Avatar 1.5 — distilled, pre-quantized (FP8 + NVFP4)

Derivative of meituan-longcat/LongCat-Video-Avatar-1.5 (MIT License, see LICENSE).

What was changed:

• the DMD distillation LoRA (lora/dmd_lora.safetensors, rank 128, alpha scale 0.5, multiplier 1.0) is merged into the weights;
• quantized linear layers are stored ready for inference: FFN (w1, w2, w3) of blocks 4–43 in NVFP4 (e2m1 values, e4m3 per-16 block scales in cuBLAS blocked layout, fp32 per-tensor scale), all other quantized layers in FP8 e4m3 with per-row scales. The source weights were the official INT8 checkpoint (base_model_int8).

NVFP4 layers require an NVIDIA Blackwell GPU (sm_100/sm_120). Layer formats are listed in base_model_okno/okno_quantization.json. The remaining components (VAE, text encoder, Whisper, vocal separator) are unchanged and should be taken from the original repositories.

Files are split into ~1 GB shards for parallel download.