Note Explore existing public work across Apple’s MLX ecosystem: three merged MLX/MLX-LM fixes, 20 model artifacts, training data, two Spaces, and technical documentation. Start with the contribution records, try Wisp, or follow a pinned setup and evaluation guide.

Note Reusable code, specialist training material, and pruning calibration from GLM Demolition. Includes a seven-example MLX code sample. The default Hub viewer indexes 87,586 rows; the full JSONL repository has a broader scope. See the card for provenance and measured limits.

Note Try Wisp in your browser: free CPU inference, fill-in-the-middle and ordinary completion, editable examples, and explicit limits. Runs the standard trunk; no MTP acceleration. Generated code is displayed without execution.

philipjohnbasile/wisp-coder-110m

Text Generation • 0.1B • Updated 17 days ago • 982 • 2

Note My small code-completion research model, trained with fill-in-the-middle and native MTP. Standard Transformers runs the 100.7M trunk; the MTP sidecar needs its explicit MLX runtime. Includes null and adverse evaluation results.

philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX

Text Generation • 35B • Updated 17 days ago • 2.03k • 5

Note My MTPLX/MLX conversion of Ornith AI's 35B-A3B model: mixed 4/8-bit body, BF16 MTP, and preserved vision weights. About 22.09 GB of weights. Recorded Forge results are limited; see runtime and licensing notes before use.

philipjohnbasile/Qwen3.6-27B-Fable-Fusion-711-MTPLX-6bit

Image-Text-to-Text • 27B • Updated 17 days ago • 452 • 8

Note My 6-bit MLX conversion of DavidAU's Fable-Fusion-711, with a calibrated MTP sidecar and vision. About 23.6 GB of weights. M5 Max measurements include runtime versions and conditions; the card retains the corrected refusal results.