Sawfwair/Cosmos3-Super-Text2Image-4Step-MLX-4bit

Cosmos3-Super Text2Image 4-Step MLX 4-bit

This repository is a native MLX conversion of nvidia/Cosmos3-Super-Text2Image-4Step, pinned at revision aa0d5a57b7b045d68daa60fbacd84ec723c7cb7b.

The 64-billion-parameter mixture-of-transformers backbone uses 4-bit affine group quantization with a group size of 64. The VAE and normalization, projection, and time-embedding tensors retain their source precision. The unused sound projections and language-model head are omitted because this artifact supports text-to-image generation only.

Use with mere.run

Use the native Cosmos3 command and pass the downloaded repository as a local model root:

mere.run video cosmos3 \
  "A glass observatory above a bioluminescent ocean at blue hour" \
  --mode text-to-image \
  --model /path/to/Cosmos3-Super-Text2Image-4Step-MLX-4bit \
  --width 768 \
  --height 768 \
  --seed 42 \
  --output cosmos3-super.png

The distilled checkpoint has a fixed four-step stochastic schedule and does not use classifier-free guidance. mere.run selects four steps and a guidance scale of 1 automatically for this artifact. Flow-shift and scheduler overrides are not supported, and negative prompts are ignored.

This native path does not bundle or run NVIDIA's separate Cosmos guardrail. Add prompt and output safety checks appropriate to your application before showing generated images to users.

Conversion and provenance

The conversion is deterministic at the tensor level and was produced by scripts/model-conversion/convert_cosmos3_super_t2i_mlx.py in the mere.run repository. See CONVERSION.json for the converter digest, package versions, hardware record, tensor counts, and quantization settings. See SOURCE_MANIFEST.json for the pinned source inventory and SHA256SUMS for the published artifact checksums.

The following source tensors are intentionally omitted:

• lm_head.weight
• audio_modality_embed
• audio_proj_in.*
• audio_proj_out.*

The upstream model card is retained as UPSTREAM_MODEL_CARD.md.

Validation status

Tensor inventory, configuration decoding, native module construction, and the published distilled scheduler are covered by automated tests. Conversion does not by itself establish visual-quality parity with NVIDIA's BF16 checkpoint. Record device, prompt, seed, output digest, peak memory, and elapsed time when qualifying a generated image.

Native qualification

The published artifact completed an end-to-end 768x768 native generation on an Apple M4 Max MacBook Pro with 128 GB of unified memory and macOS 26.5.2. The command used a debug build of mere.run, four steps, seed 42, and this prompt:

A glass observatory above a bioluminescent ocean at blue hour

Measurement	Result
Wall time	60.85 seconds
Maximum resident memory	38,657,916,928 bytes
Peak memory footprint	53,394,599,464 bytes
Swaps	0
Output SHA-256	006a73d23850c1ab6ae9c05e5521109a8a2678f69e42fe362cd3a837d77c7495

This is native runtime smoke evidence, not a BF16 parity or broad visual-quality evaluation.

License

This model is distributed under the NVIDIA Open Model Development and Weight License 1.1. The artifact includes an archived copy in OPENMDW-1.1-LICENSE.html. Review the license and the upstream use restrictions before downloading or redistributing the weights.