gajesh/LTX-2.3-mlx-fp16

LTX-2.3 22B Distilled (MLX, float16)

MLX-optimized weights for Lightricks/LTX-2.3 video generation model, converted for Apple Silicon.

Model Details

• Model: LTX-2.3 22B Distilled (joint audio-video diffusion transformer)
• Format: MLX safetensors (float16)
• Total Size: ~66GB (transformer: 39GB, text encoder: 22GB, VAE: 1.7GB, upsampler: 1GB)
• Minimum RAM: 64GB Apple Silicon Mac
• Original: Lightricks/LTX-2.3

Also available in 4-bit quantized (35GB, fits 32GB Macs).

Architecture

Component	Parameters	Size (fp16)
DiT Transformer (48 layers)	22B	39 GB
Gemma 3 12B (text encoder)	12B	22 GB
Video VAE Decoder	~500M	777 MB
Spatial Upsampler	~500M	950 MB
Audio VAE + Vocoder	~500M	348 MB

Benchmarks (M4 Max 128GB)

Head-to-head comparison on the same machine, same config (576x1024, 121 frames, 5s video):

	PyTorch MPS (BF16)	MLX FP16 (Video+Audio)	MLX FP16 (Video-only)
Stage 1 (8 steps, half-res)	66.7s	67.1s	51.2s
Stage 2 (3 steps, full-res)	157.9s	139.2s	101.3s
Total denoising	264.0s	206.2s	152.5s
Speedup	1.0x	1.3x	1.7x
Peak memory	>60 GB	46.8 GB	34.2 GB

Additional configs:

Config	Resolution	Frames	Time	Memory
FP16 video-only	512x512	49 (2s)	35.1s	34.2 GB
FP16 video-only	320x512	25 (1s)	10.4s	31.3 GB

Key Advantages

• 1.7x faster than PyTorch MPS in video-only mode
• 42% less memory (34 GB vs 60+ GB) -- runs on 64GB Macs where PyTorch can't
• Fused Metal attention via mx.fast.scaled_dot_product_attention
• Unified memory -- no CPU/GPU transfer overhead

Usage

pip install mlx-ltx

from mlx_ltx.pipeline import DistilledPipeline, save_video

pipeline = DistilledPipeline("path/to/mlx-weights")
video = pipeline(
    prompt="A golden retriever playing piano in a concert hall",
    height=576, width=1024, num_frames=121,
    seed=42,
)
save_video(video, "output.mp4", fps=24.0)

Conversion

Converted from the original PyTorch checkpoint using mlx-ltx:

mlx-ltx-convert \
  --checkpoint /path/to/ltx-2.3-22b-distilled.safetensors \
  --gemma-root /path/to/gemma-3-12b-it/ \
  --output-dir ./mlx-weights/

License

This model conversion follows the LTX-2 Community License.

Acknowledgments

• Lightricks for the original LTX-2.3 model
• MLX by Apple for the framework
• Acelogic/LTX-2-MLX for reference implementation patterns