Instructions to use josand/MLX-Reason-CT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use josand/MLX-Reason-CT with MLX:

# Make sure mlx-vlm is installed
# pip install --upgrade mlx-vlm

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load the model
model, processor = load("josand/MLX-Reason-CT")
config = load_config("josand/MLX-Reason-CT")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use josand/MLX-Reason-CT with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "josand/MLX-Reason-CT"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "josand/MLX-Reason-CT"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use josand/MLX-Reason-CT with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "josand/MLX-Reason-CT"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default josand/MLX-Reason-CT

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use josand/MLX-Reason-CT with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "josand/MLX-Reason-CT"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "josand/MLX-Reason-CT" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

josand commited on 17 days ago

Commit

eef8b72

verified ·

1 Parent(s): e15558a

Simplify model card

Browse files

Files changed (1) hide show

README.md +11 -56

README.md CHANGED Viewed

@@ -15,64 +15,19 @@ tags:
 # NV-Reason-CT MLX
-Private **FP32** converted weights for the standalone
-[NV-Reason-CT MLX code](https://github.com/sandovaljoseph/NV-Reason-CT-MLX).
-This is an unofficial port by Alfredo and Joseph Sandoval.
-These weights **losslessly expand NVIDIA's original BF16 parameters into FP32**.
-No retraining, fine-tuning, or additional parameter precision is claimed.
-Inference and cache arithmetic in this release use FP32. This artifact is not
-original BF16 inference and does not contain a separately trained FP32 checkpoint.
-Upstream: [NVIDIA NV-Reason-CT](https://huggingface.co/nvidia/NV-Reason-CT/tree/386b93e034983f6c1fc841a43833a1b6a0cd9c13).
-Revision: `386b93e034983f6c1fc841a43833a1b6a0cd9c13`.
-Original checkpoint SHA256:
-`655c15e59dfa1957f7d5f475b1702d689fb7107160f21256028391bb5f6e6ba5`.
-The bundle has 35 FP32 safetensors shards, 17,411,301,664 bytes, covering
-804 active tensors. Every converted storage pattern was independently checked
-against the pinned original checkpoint. The inactive 2D tower, unused mask token,
-and duplicated tied LM head are excluded; exclusions are listed in `mlx/manifest.json`.
-## Usage
-Install the standalone code on macOS Apple Silicon with Python 3.12. Authenticate
-an account with access to this private model repository. Pin this repository's
-commit rather than an unversioned moving download.
-```bash
-uv run nv-reason-ct-mlx download --revision MODEL_COMMIT --model-dir models/NV-Reason-CT-MLX
-uv run nv-reason-ct-mlx verify --model-dir models/NV-Reason-CT-MLX
-uv run nv-reason-ct-mlx report \
-  --input /path/to/ct.nii.gz \
-  --model-dir models/NV-Reason-CT-MLX \
-  --anatomy-region chest \
-  --prompt "write a structured chest CT report" \
-  --max-new-tokens 512 \
-  --output-dir outputs/chest
-```
-All runtime metadata and shard hashes are verified before inference. The original
-PyTorch checkpoint is not needed. This layout is for the standalone port and is
-not loaded with Transformers AutoModel or the general MLX-LM loader.
-## Validation limits
-Historical FP32 synthetic engineering gates cover four cases and exact generated
-tokens: `fp32-engineering-qualification.json`. Clean-wheel CPU image/text parity,
-full tensor conversion integrity and package checks are in `standalone-verification.json`.
-Fresh standalone learned inference remains pending while the shared Metal device
-is occupied by another qualification run. No clinical accuracy is established.
-The model is for research and education, not clinical diagnosis or treatment decisions.
-Outputs require human review and are not ground truth. Preserve the upstream
-limitations and dataset/training descriptions in [the original model card](upstream_model_card.md).
-## License
-NVIDIA model materials, including these converted parameters, retain **OpenMDW-1.1**
-([LICENSE](LICENSE)), Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
-No separate permission from NVIDIA is implied or required beyond complying with
-those terms. The standalone code has its own Apache-2.0 and retained third-party terms.
-Underlying Qwen3.5 materials retain their applicable Apache-2.0 terms; the license
-text is included in `APACHE-2.0.txt`.

 # NV-Reason-CT MLX
+FP32 [NV-Reason-CT](https://huggingface.co/nvidia/NV-Reason-CT) weights for Apple Silicon.
+An unofficial port by Alfredo and Joseph Sandoval.
+**[Install and run →](https://github.com/sandovaljoseph/NV-Reason-CT-MLX#quick-start)**
+35 shards · 17.4 GB · exact BF16-to-FP32 expansion · no retraining.
+All weight hashes and four CPU preprocessing cases are verified.
+Fresh standalone GPU inference is pending.
+[Verification](standalone-verification.json) · [FP32 evidence](fp32-engineering-qualification.json)
+Research and education only. Reports require human review; not for clinical decisions.
+[Upstream model card and limitations](upstream_model_card.md).
+**License:** [OpenMDW-1.1](LICENSE), Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+Underlying Qwen3.5 terms: [Apache-2.0](APACHE-2.0.txt).