This conversion is intended for local inference on Apple Silicon using MLX / MLX-VLM. It preserves the upstream model architecture and tokenizer/processor files where applicable, while quantizing supported linear weights to 4-bit for smaller memory footprint.

Attribution, license, and required notices

Base model: nvidia/Cosmos-Reason2-2B
License: NVIDIA Open Model License
Required attribution: Licensed by NVIDIA Corporation under the NVIDIA Open Model License
Required Cosmos notice: Built on NVIDIA Cosmos
This repository is a derivative conversion/quantization and is not the original NVIDIA release.

A NOTICE file is included in this repository. By using, copying, modifying, redistributing, deploying, or making this derivative model available to others, you are responsible for complying with all applicable terms, including but not limited to:

the NVIDIA Open Model License;
NVIDIA Trustworthy AI terms and any responsible-use, prohibited-use, or acceptable-use restrictions that apply to the upstream model;
export-control, sanctions, and other applicable laws and regulations;
any additional terms, notices, access requirements, or usage instructions presented on the original NVIDIA model page.

If the upstream license, notices, or model-page terms are updated, those upstream terms may impose additional or different obligations. Please review the upstream model page and license before use or redistribution.

No affiliation, sponsorship, endorsement, or trademark grant

This repository is independently prepared and published by the repository owner. It is not affiliated with, sponsored by, approved by, or endorsed by NVIDIA Corporation unless NVIDIA explicitly states otherwise.

The names "NVIDIA", "Cosmos", and other NVIDIA marks are trademarks or registered trademarks of NVIDIA Corporation in the United States and/or other jurisdictions. They are used here only for reasonable descriptive attribution and identification of the upstream base model. No trademark license or other rights in NVIDIA marks are granted by this repository.

Conversion details

Source: nvidia/Cosmos-Reason2-2B
Format: MLX / MLX-VLM
Quantization: 4-bit affine weight quantization
Group size: 32
Conversion command:

bash

python3 -m mlx_vlm.convert \
  --hf-path nvidia/Cosmos-Reason2-2B \
  --mlx-path Cosmos-Reason2-2B-Arbor-4bit-g32 \
  -q \
  --q-bits 4 \
  --q-group-size 32 \
  --trust-remote-code

Usage

Install MLX-VLM:

pip install -U mlx-vlm

Example text prompt:

bash

python -m mlx_vlm.generate \
  --model masato25/Cosmos-Reason2-2B-Arbor-4bit-g32 \
  --prompt "Explain physical common sense reasoning in 3 bullets." \
  --max-tokens 256 \
  --temperature 0.0

Example image prompt:

bash

python -m mlx_vlm.generate \
  --model masato25/Cosmos-Reason2-2B-Arbor-4bit-g32 \
  --image /path/to/image.jpg \
  --prompt "Describe the scene and reason about what may happen next." \
  --max-tokens 256 \
  --temperature 0.0

Intended use

This quantized derivative is intended for experimentation, prototyping, and deployment scenarios where the upstream license permits use and where Apple Silicon local inference is desirable.

You should evaluate whether your intended use is permitted under the upstream NVIDIA Open Model License and related terms. This repository does not expand, waive, or modify any upstream restrictions.

Limitations and safety

Quantization can change output quality, numerical behavior, robustness, and safety characteristics compared with the original model. This repository does not claim improved accuracy, safety, bias mitigation, alignment, or suitability for any particular purpose over the upstream NVIDIA model.

Model outputs may be inaccurate, unsafe, biased, offensive, incomplete, or otherwise unsuitable for your use case. Do not rely on the model as the sole source of truth for medical, legal, financial, safety-critical, or other high-stakes decisions. Evaluate the model for your own task, domain, jurisdiction, and risk profile before deployment or redistribution.

Disclaimer

This derivative model is provided as-is and without warranties or conditions of any kind, express or implied, including without limitation warranties of merchantability, fitness for a particular purpose, title, non-infringement, accuracy, availability, or error-free operation.

To the maximum extent permitted by applicable law, the repository owner is not liable for any direct, indirect, incidental, special, consequential, exemplary, punitive, or other damages arising from or related to use of this repository, the derivative model, or model outputs.

Nothing in this README is legal advice. You are responsible for reviewing and complying with the applicable license terms and laws.