Apple MLXMerged
Correct results beyond 32K rows
A Metal kernel fix for row-count overflow in sorted quantized matrix multiplication.
#3922Aug 26, 2026Open source
Make it work.
Share what helps.
I contribute fixes upstream and publish the models, code, and research records that come out of working with AI on Apple Silicon.
Apple MLX ecosystem
Contributions in the foundation.
My connection to Apple’s MLX work is through public, reviewed contributions to MLX and MLX-LM: correctness fixes in Metal kernels and model conversion.
Apple MLXMerged
Repair quantized loader indexing
Corrected scale and bias indexing for small groups, with upstream dispatch changes in the merged patch.
#4202Aug 12, 2026Apple MLX-LMMerged
Keep converted Qwen weights correct
Prevented a second normalization shift when converted checkpoints retained MTP tensors.
#1623Aug 18, 2026Beyond the patch
A connected body of work.
Model + demo
Wisp Coder ↗
A small code model trained from scratch with MLX. Explore its fill-in-the-middle experiments, native MTP head, and public demo.
Runtime + systemsLocal inference ↗
From MTPLX contribution threads to iliria’s C/Metal engine and SSD-backed expert streaming.
Source + documentationHy3 demolition MLX ↗
Model work with architecture notes, restoration instructions, and an explicit 64 GB feasibility analysis.
Training dataGLM-5.2 demolition data ↗
A public dataset with a documented source revision and sample inspection. Read the card for coverage and format details.
Models, datasets, Spaces, and docs
The public hub connects releases to source code, training records, and live demos.
Project and pull-request links are the source of truth for release status and technical details.