Kiwi-Edit 5B (instruct-only), VideoX-Fun layout
Frozen base components shared by the four Video-Editor-MoE models (explicit-moe, mlp-moe, videoeditor-14b, videoeditor-56b), converted from the public Kiwi-Edit release to the layout used by the Video-Editor-MoE code base.
huggingface-cli download Massyzs/kiwi-edit-5b-instruct-only-videoxfun --local-dir <BASE_DIR>/ckpt/kiwi-edit-instruct-only-videoxfun
Files
| File | Size | Content |
|---|---|---|
config.json | 423 B | DiT config (Wan2.2-TI2V-5B architecture: 30 blocks, dim 3072, FFN 14336, 24 heads, 48 latent channels) |
diffusion_pytorch_model.safetensors | 10.0 GB | Kiwi-Edit instruct-only DiT, 825 tensors, bf16, VideoX-Fun key names |
moe_expert_init.safetensors | 34.8 MB | mllm.connetor.* (MLLM-to-DiT connector), mllm.video_queries (learnable query tokens), dit.vae_condition.* (source-video patch embedding) |
Wan2.2_VAE.pth | 2.8 GB | Wan2.2 video VAE |
qwen/ | 7.5 GB | Kiwi-Edit MLLM encoder (Qwen2.5-VL-3B architecture, fine-tuned by Kiwi-Edit) with tokenizer and processor configs; loadable with from_pretrained |
SHA256SUMS lists the checksum of every file (sha256sum -c SHA256SUMS).
Provenance
| File | Original | Relation |
|---|---|---|
diffusion_pytorch_model.safetensors | transformer/ of linyq/kiwi-edit-5b-instruct-only-diffusers | same tensors, keys renamed from diffusers to VideoX-Fun naming |
qwen/ | mllm_encoder/ + processor/ of the same repo | weight shards byte-identical; note this is the Kiwi-Edit fine-tuned encoder, not stock Qwen/Qwen2.5-VL-3B-Instruct |
moe_expert_init.safetensors | mllm_encoder/diffusion_pytorch_model.safetensors + source_embedder/ of the same repo | same tensors, renamed; image query tokens dropped (only the video path is used) |
Wan2.2_VAE.pth | Wan-AI/Wan2.2-TI2V-5B | byte-identical |
License
Kiwi-Edit is released under the MIT license by its authors (paper 2603.02175,
code showlab/Kiwi-Edit). Wan2.2_VAE.pth is part of Wan2.2, released by Alibaba under Apache-2.0.
Please cite Kiwi-Edit and Wan2.2 if you use these weights.