MLX-Reason-CT 0.2.2

Fix public installation by pinning the public Apache-2.0 medmlx-core v0.1.2 release. Existing tags remain unchanged. Quick-start instructions now reference v0.2.2.

Validation on macOS arm64 / Python 3.12.12 / MLX 0.32.3 Metal:

• 95 tests passed; formatting, lock consistency, wheel/source builds and metadata checks passed.
• Fresh wheel installation succeeded with Git credentials, user configuration and caches disabled; dependency compatibility and CLI checks passed.
• The installed wheel executed the full FP32 model with the pinned 35-shard bundle on a synthetic HU NIfTI chest volume. It generated a nonempty two-token response terminating at EOS and wrote report.txt, model_response.json and run.json (45.97 seconds, 22.75 GB peak MLX memory). This is an inference smoke check, not report accuracy, full-model parity or clinical validation.
• Publication scan found no blockers; the only warning was a generated token ID, not a credential.

Known pre-existing QA backlog: make qa stops on 144 Ruff findings; strict Pyright reports 161 errors. These remain visible and are not suppressed by this packaging release.

MLX-Reason-CT 0.2.1

Adds explicit float32, bfloat16 and bfloat16_fp32 precision profiles for local 3D CT generation. FP32 remains the default. Includes the MedMLX runner, explicit overwrite controls, and macOS arm64 Metal runtime checks.

Install the 0.2.1 wheel:

uv tool install --python 3.12 \
  https://github.com/MedMLX/MLX-Reason-CT/releases/download/v0.2.1/mlx_reason_ct-0.2.1-py3-none-any.whl

The usage guide and model card use the same runtime version and pinned model bundle.

Validation: make env, make qa (95 tests, Ruff, strict Pyright, 100% type completeness), make build, and clean-wheel checks for CLI profiles, MedMLX discovery and packaged Metal arithmetic. See verification scope for the bounds of these checks.

MLX-Reason-CT 0.2.0

Source release for the existing v0.2.0 tag at commit 3a015e2.

Package: mlx-reason-ct==0.2.0. Python requirement: >=3.12,<3.13.

Install this tagged snapshot:

uv pip install "mlx-reason-ct @ git+https://github.com/MedMLX/MLX-Reason-CT.git@v0.2.0"

Tagged source and usage. License and model-use terms are those included in that tagged snapshot.

Changes since v0.1.1.

This is a historical release; v0.2.1 is newer.

This entry reconciles the release catalog with a previously published tag. It does not include later commits on main, add binary or model-weight assets, or assert new inference or numerical validation.

MLX-Reason-CT 0.1.1

Update project and documentation links after moving the GitHub repository to MedMLX. Runtime code, dependencies, and model weights are unchanged.

Validation: 26 tests passed; Ruff, strict Pyright, type completeness (100%), wheel/sdist build, and Twine checks passed.

MLX-Reason-CT 0.1.0

Native MLX port of NVIDIA NV-Reason-CT for local 3D CT reasoning and report generation on Apple Silicon.

• Chest and abdomen CT reports and question answering from local NIfTI volumes.
• FP32 inference on the Apple Silicon GPU, with a command-line interface and typed Python API.
• Portable CPU preprocessing, checkpoint conversion, and bundle verification.

Install into a Python 3.12 environment:

pip install mlx-reason-ct

Or install the command-line tool in an isolated environment:

uv tool install --python 3.12 mlx-reason-ct

Download the model weights at revision e15558ae30c8ad6c25c0bfcce7467cc72cd0b2f2; follow the quick start and usage guide.

Inference requires macOS on Apple Silicon. Weights occupy 17.4 GB; measured peak MLX memory use is about 22.7 GB, with additional memory needed for preprocessing and macOS.

The wheel and source archive are attached with SHA-256 checksums. Code is Apache-2.0; model weights retain NVIDIA's OpenMDW-1.1 terms and notices. Intended for research and education; generated outputs require human review.