LongCat-Video-Avatar-1.5-bf16-dmd-merged (MLX)

Apple MLX bf16 weights for LongCat-Video-Avatar-1.5 — Meituan's audio-driven video diffusion model — with the DMD step-distillation LoRA pre-merged into the DiT weights. Recommended variant for inference: produces the same outputs as the base + LoRA combination but in 8 sampling steps with no LoRA loading at runtime.

TL;DR


Architecture	Wan 2.1 VAE + umT5-XXL + Whisper-Large-v3 + 48-block Avatar DiT (with DMD LoRA pre-merged)
Params	~13.6 B DiT + ~11 B umT5 + ~0.6 B Whisper encoder + 0.5 B VAE
Format	bf16, sharded safetensors (HF-style per-component subdirs)
Disk	~43 GB
Hardware	Apple Silicon M-series, 64 GB+ unified memory recommended for 480p
Inference	8-step DMD distilled, FlowMatchEulerDiscreteScheduler with shift=7.0
License	MIT (matches upstream Meituan)

Quick start

# 1. Pull weights (~43 GB)
hf download mlx-community/LongCat-Video-Avatar-1.5-bf16-dmd-merged \
    --local-dir ./weights

# 2. Set up inference (Python 3.12)
git clone https://github.com/xocialize/longcat-avatar-mlx
cd longcat-avatar-mlx
python3.12 -m venv .venv
.venv/bin/pip install -e ".[parity]"
.venv/bin/pip install librosa Pillow imageio imageio-ffmpeg

# 3. Run end-to-end
.venv/bin/python scripts/run_inference.py \
    --weights ./weights/.. \
    --variant merged \
    --num-frames 93 \
    --out output.mp4

Programmatic usage:

import json
import pathlib
import mlx.core as mx

from longcat_video_avatar.pipeline_mlx import LongCatAvatarPipeline, PipelineConfig
from longcat_video_avatar.models.autoencoder_kl_wan import AutoencoderKLWan
from longcat_video_avatar.models.avatar.longcat_video_dit_avatar import (
    LongCatVideoAvatarTransformer3DModel,
)
from longcat_video_avatar.models.umt5 import UMT5EncoderModel
from longcat_video_avatar.models.whisper import WhisperEncoder

W = pathlib.Path("./weights")

vae = AutoencoderKLWan.from_config(json.loads((W/"vae/config.json").read_text()))
vae.load_weights(str(W/"vae/diffusion_pytorch_model.safetensors"), strict=False)

# (Load umT5/Whisper/DiT similarly via from_config + load_weights; see
# scripts/run_inference.py:build_pipeline() for the sharded-load helper.)

pipeline = LongCatAvatarPipeline(vae=vae, text_encoder=umt5,
                                 audio_encoder=whisper, dit=dit,
                                 config=PipelineConfig())
video = pipeline(image=ref_image_chw, audio_mel=mel_features,
                 text_embeds=text_embeds, text_mask=text_mask,
                 uncond_embeds=neg_embeds, uncond_mask=neg_mask)

Variants

Variant	DiT dtype	Disk	Sampling	Best for
bf16-dmd-merged (this card)	bf16	43 GB	8-step	64 GB+ Macs, recommended baseline
bf16	bf16 (+ separate LoRA)	46 GB	8-step or 50-step	runtime-merge / multi-strength experiments
q4-dmd-merged	4-bit quantized	24 GB	8-step	32–48 GB Macs, comparable speed to bf16
q8-dmd-merged	8-bit quantized	31 GB	8-step	middle ground RAM / quality

Performance

Tested on Apple M5 Max (128 GB unified memory):

Resolution	Frames	Wall clock	ms/frame
256 × 432	29	~105 s	~3.6 s
480 × 832	93	TBD (estimate ~20–30 min)	—

Peak memory at 480 × 832: ~50 GB unified. The VAE attention layers are routed to a CPU stream by default (see CLAUDE.md L10) to recover strict fp32 precision; net perf hit is negligible.

Layout

LongCat-Video-Avatar-1.5-bf16-dmd-merged/
├── README.md                           # this file
├── pipeline_config.json
├── vae/
│   ├── config.json
│   └── diffusion_pytorch_model.safetensors        # ~254 MB (bf16)
├── text_encoder/
│   ├── config.json
│   ├── model.safetensors.index.json
│   └── model-{00001,00002,00003}-of-00003.safetensors   # ~11 GB total (bf16)
├── audio_encoder/                                  # Whisper-large-v3 ENCODER only
│   ├── config.json
│   └── model.safetensors                           # ~1.3 GB (bf16)
├── dit/                                            # DMD LoRA pre-merged
│   ├── config.json
│   ├── diffusion_pytorch_model.safetensors.index.json
│   └── diffusion_pytorch_model-{00001..00007}-of-00007.safetensors   # ~33 GB total
├── scheduler/
│   └── scheduler_config.json                       # FlowMatchEuler, shift=7.0
└── tokenizer/                                       # umT5 tokenizer files
    ├── tokenizer.json
    ├── tokenizer_config.json
    └── special_tokens_map.json

Source weights

Provenance, in case you want to verify or re-derive these weights:

Subdir	Source	Conversion
`vae/`	`meituan-longcat/LongCat-Video/vae/`	Conv3d weight transpose `(O,I,T,H,W)→(O,T,H,W,I)`; dtype passthrough
`text_encoder/`	`meituan-longcat/LongCat-Video/text_encoder/`	HF verbose → mlx-compact key rename (`shared`→`token_embedding`, `encoder.block.{B}.layer.{0,1}.…`→`blocks.{B}.{attn,ffn,norm{1,2}}.…`); dtype cast to bf16
`audio_encoder/`	`meituan-longcat/LongCat-Video-Avatar-1.5/whisper-large-v3/model.safetensors`	`model.encoder.` prefix strip; Conv1d weight transpose; encoder-only
`dit/`	`meituan-longcat/LongCat-Video-Avatar-1.5/base_model/` + `…/lora/dmd_lora.safetensors`	passthrough names; DMD LoRA merged into 336 modules (7 LoRA targets × 48 blocks); adaLN_modulation weights kept at fp32
`scheduler/`	`meituan-longcat/LongCat-Video-Avatar-1.5/scheduler/`	verbatim copy
`tokenizer/`	`meituan-longcat/LongCat-Video/tokenizer/`	verbatim copy

Conversion recipe: recipes/convert_longcat_avatar.py in the companion GitHub repo. Run with --variant merged --out <dir> to reproduce these weights from Meituan's PT sources.

Numerical conventions preserved from upstream

fp32 internal compute for RMSNorm / LayerNorm / AdaLN modulation (Meituan's _FP32 suffix convention) — adaLN_modulation and gamma weights are stored fp32 to defend this.
Negative velocity flip before scheduler step (noise_pred = -noise_pred) — Meituan's DiT outputs -v; we flip to +v for FlowMatchEulerDiscreteScheduler.step.
3-pass disentangled CFG combiner: uncond + s_t·(cond − uncond_text) + s_a·(uncond_text − uncond). Defaults: s_t = s_a = 4.0 (matches PT DMD distillation).
DMD distilled sigma schedule — 8 sigmas spanning [1.0, 0.124] computed by guidance.get_dmd_distilled_sigmas. The pipeline overwrites the trailing sentinel sigma (mlx-arsenal appends 1.0; we replace with 0.0 to actually denoise to clean at the final step — see CLAUDE.md L17).

License

MIT. Matches upstream Meituan LongCat-Video license. Adapted code from Blaizzy/mlx-video (MIT) for some MLX op primitives; full attribution in LICENSE.

Citation

@misc{longcat-avatar-mlx,
  title  = {longcat-avatar-mlx: Apple MLX port of LongCat-Video-Avatar-1.5},
  author = {xocialize},
  year   = {2026},
  url    = {https://github.com/xocialize/longcat-avatar-mlx},
}

@techreport{meituan2026longcat,
  title       = {LongCat-Video-Avatar 1.5 Technical Report},
  author      = {Meituan LongCat Team},
  institution = {Meituan},
  year        = {2026},
  url         = {https://github.com/meituan-longcat/LongCat-Video},
}