Apple MLX weights for LongCat-Video-Avatar 1.6 (Meituan, released 15 Nov 2026): a portrait (or a few frames of real footage) plus a voice track in, a talking video of any length out. Runs locally on an Apple Silicon Mac with 64 GB of unified memory.
1.6 keeps the 1.5 architecture with a retrained base DiT and two LoRAs: DMD (step distillation) and
OPSD (on-policy distillation). Meituan's recommended recipe samples 8 steps with DMD for the first 4
and OPSD for the last 4 (--mix_steps 4 --mix_alpha 1.0); the code below applies both at runtime and
switches them exactly at step 4.
| folder | what | source |
|---|---|---|
dit/ |
1.6 base DiT (48 blocks, 13.6B) converted to this port's layout: conv weights channels-last, AdaLN modulation fp32, the rest bf16 (bit-exact passthrough of Meituan's bf16) | meituan-longcat/LongCat-Video-Avatar-1.6/base_model |
lora/dmd_loraA.safetensors, lora/opsd_loraB.safetensors |
the two 1.6 LoRAs, unchanged (rank 128, Meituan key format) | loraA/, loraB/ |
vae/, text_encoder/, audio_encoder/, tokenizer/ |
Wan 2.1 VAE, umT5-XXL, Whisper-large-v3 encoder: the same as 1.5 | mlx-community/LongCat-Video-Avatar-1.5-bf16 |
51.5 GB on disk.
The inference code is in xocialize/longcat-avatar-mlx (PR #8 adds 1.6, long videos and LoRA training; until it is merged, use the branch below):
git clone -b feat/avatar-1.6-continuation-lora https://github.com/fedeinfe/longcat-avatar-mlx
cd longcat-avatar-mlx
python3.12 -m venv .venv && .venv/bin/pip install -e ".[parity]" librosa pyloudnorm pillow ftfy
hf download fedeinfe/LongCat-Video-Avatar-1.6-MLX-bf16 --local-dir weights/LongCat-Video-Avatar-1.6-bf16
.venv/bin/python scripts/generate.py --weights weights/LongCat-Video-Avatar-1.6-bf16 \
--image portrait.png --audio voice.wav \
--prompt "A young man with curly hair sits in his apartment and talks to the camera ..." \
--out talking.mp4
A long voice track comes out as one continuous shot: 93-frame segments at 25 fps, each continuing from the last 13 frames of the previous one plus the reference portrait (Meituan's video continuation).
| 480p bucket (704x544 for 4:5) | 1120x832 | |
|---|---|---|
| s / step, first segment | ~55 | ~320 |
| one 93-frame segment (3.7 s), 8 steps + decode | ~8 min | ~45 min |
| peak memory | ~47 GB (bf16 VAE) | ~56 GB (fp32 VAE, --vae-tile 64) |
Continuation segments are ~20% cheaper thanks to the conditioning KV cache. Generation is 25 fps by design: feeding audio at 30 fps desynchronizes the lips (SyncNet confidence drops from ~7 to ~2); interpolate afterwards (e.g. RIFE) if you need 30 fps.
Lip sync on our tests (SyncNet LSE-C, higher is better): 6.8-7.4 for 1.6, 6.5 for 1.5, against 7.05 for real footage of the same person. The compiled path (default) runs ~45 s/step at 480p.
--audio-scale 0.7 --audio-steps 3 costs 1.4x).training/train_lora.py in the PR (rank 32, ~49.6 GB peak and ~65 s/step at 49 frames 640x512).MIT, like the original. Model by the Meituan LongCat Team; MLX port by xocialize; 1.6 conversion, long-video pipeline and LoRA training by fedeinfe.
dit/diffusion_pytorch_model-00003-of-00006.safetensors 5.8 GBdit/diffusion_pytorch_model-00002-of-00006.safetensors 5.7 GBdit/diffusion_pytorch_model-00005-of-00006.safetensors 5.7 GBdit/diffusion_pytorch_model-00004-of-00006.safetensors 5.6 GBdit/diffusion_pytorch_model-00006-of-00006.safetensors 5.5 GBtext_encoder/model-00001-of-00003.safetensors 5.4 GBdit/diffusion_pytorch_model-00001-of-00006.safetensors 5.3 GBtext_encoder/model-00002-of-00003.safetensors 5.3 GBlora/dmd_loraA.safetensors 2.5 GBlora/opsd_loraB.safetensors 2.5 GBaudio_encoder/model.safetensors 1.3 GBtext_encoder/model-00003-of-00003.safetensors 671 MBvae/diffusion_pytorch_model.safetensors 254 MB