Part of the LongCat-Video-Avatar 1.5 — MLX collection.

LongCat-Video-Avatar-1.5-q4-dmd-merged (MLX)

4-bit quantized variant of mlx-community/LongCat-Video-Avatar-1.5-bf16-dmd-merged. Same model, same DMD pre-merge, same 8-step inference path — just with the DiT Linears quantized to 4-bit via mlx.nn.quantize for smaller-RAM Macs.

DiT	4-bit quantized (group_size=64, skip final_layer.linear + embedders + AdaLN)
DiT shards	~11 GB (3 shards)
umT5 / Whisper / VAE	bf16 (unchanged from the bf16-dmd-merged variant)
Total disk	~24 GB
Min unified memory	~24 GB
Inference	8-step DMD distilled (unchanged)
License	MIT

Performance

Measured on Apple M5 Max (128 GB unified memory), 256 × 432 × 29 frames, 8-step DMD sampling:

Variant	Wall clock	ms/frame
bf16-dmd-merged	~105 s	~3.6 s
q4-dmd-merged	~102 s	~3.5 s
q8-dmd-merged	~151 s	~5.2 s

q4 is bandwidth-bound (matches bf16 throughput); q8 currently runs slower on M5's quantized matmul kernels but uses ~half the DiT disk vs bf16. Pick the variant by RAM budget, not speed.

Loading

The runtime pipeline (longcat_video_avatar.pipeline_mlx.LongCatAvatarPipeline) auto-detects the quantization block in dit/config.json and applies mlx.nn.quantize before loading the quantized weights. No user-facing API change vs. the bf16 variant.

hf download mlx-community/LongCat-Video-Avatar-1.5-q4-dmd-merged \
    --local-dir ./weights
.venv/bin/python scripts/run_inference.py \
    --weights ./weights/.. \
    --variant q4-merged \
    --num-frames 93 \
    --out output.mp4

Source

Quantized from the bf16-dmd-merged variant via recipes/convert_longcat_avatar.py. Run with --variant q4-merged --out <dir> to reproduce from Meituan's PT sources.

See the bf16-dmd-merged card for full architecture details, citation, and the non-quantized variant.