JuliaML/Cosmos3-Super-Text2Image-4Step-INT8-G64-BF16

Cosmos3 Super Text-to-Image 4-Step — INT8 G64 BF16

Independent transformer-only quantization of nvidia/Cosmos3-Super-Text2Image-4Step. This is not an official NVIDIA release.

Base revision: 0573a4b26b8e15d13d416e51f4680c8bc8b8c33d.

Runtime, verification and benchmark code: gtrg55/cosmos3-quant-mlx-cuda.

Format

Property	Value
Quantization	weight-only signed symmetric INT8 (W8A16)
Group size	64
Compute / activations	BF16
BF16 transformer I/O	proj_in, proj_out
Quantized matrices	898
Transformer safetensors	64.77 GiB
SVD / Hadamard / dynamic allocation	disabled
Base sampler	original distilled 4-step schedule, guidance 1.0

The repository contains the quantized transformer. Tokenizer, VAE and scheduler are loaded from the original NVIDIA model.

Validation status

• Apple MLX / M4 Max: validated end-to-end at 1024×1024.
• Checkpoint packing: exact sample round-trip verified.
• CUDA: loader and test command are supplied in the companion code repo; publish a CUDA performance claim only after completing the included RunPod test.

Reference M4 Max result: 70.54 seconds generation, 67.63 GiB MLX peak memory. Hardware, thermals and software versions affect timings.

Download and run on Apple Silicon

hf download nvidia/Cosmos3-Super-Text2Image-4Step \
  --revision 0573a4b26b8e15d13d416e51f4680c8bc8b8c33d \
  --local-dir models/Cosmos3-Super-Text2Image-4Step

hf download JuliaML/Cosmos3-Super-Text2Image-4Step-INT8-G64-BF16 \
  --local-dir checkpoints/Cosmos3-INT8

python mlx/run_pipeline.py \
  --model models/Cosmos3-Super-Text2Image-4Step \
  --checkpoint checkpoints/Cosmos3-INT8/transformer \
  --prompt prompts/winter-spruce-tall-1024.json \
  --latents benchmarks/latents/cpu-seed1143-1024x1024-bf16.safetensors \
  --output outputs/int8.png \
  --height 1024 --width 1024 --seed 1143

Runtime code and pinned requirements are maintained in the companion GitHub repository.

Reproducibility assets

examples/ contains the exact structured prompt, fixed BF16 latent, generation configuration, output image and run report. Important hashes:

• prompt SHA-256: 3271fe4b38b1882cde4d3c87a4d4ae260c568cf4ab2d35c5d0f32bb12b1ba936
• latent SHA-256: 6504206b2452a23680cac73d5763e3796260f7582f40ff59e6613cdb33f0aca6
• reference PNG SHA-256: 22e3724df99afa953699f46c6529c042b4d37e505d39fef99da6c03c1f59b8b9

Limitations

• INT8 remains extremely close to the affine W8 control in the included test, but this one prompt is not a comprehensive evaluation.
• This is weight-only quantization, not integer activation inference.
• Safety behavior is inherited from the base pipeline and the user's runtime configuration.

License and attribution

Distributed under OpenMDW-1.1. See LICENSE and NOTICE. The original NVIDIA notices and license must remain with redistributed model materials.