MLX · Metal · Machine learning on Mac
AI built for your Mac. I contribute to Apple's MLX and MLX-LM, build local inference tools, and publish models, training data, and reproducible research.
Apple Silicon. Unified memory. Open source.
Contributing upstream
MLX is Apple’s open-source framework ↗ for machine learning on Apple Silicon. These contributions connect my local model work to the tools the wider community uses.
Fixed a row-count overflow in the sorted quantized matrix kernel that could leave output rows unwritten above 32K. A correctness fix validated on M5 Max.
View upstream contribution ↗August 26, 2026Fixed scale and bias indexing for small quantization groups. The merged change also disabled the affected non-transposed NAX dispatch while upstream reviewed it.
View upstream contribution ↗August 12, 2026Fixed a second normalization shift when already-converted Qwen checkpoints retained MTP tensors. The regression covers dense and mixture-of-experts models.
View upstream contribution ↗August 18, 2026Merged contributions, verified September 10, 2026. The linked pull requests preserve the implementation, review, and validation scope.
From frameworks to applications
Runtime development, model engineering, and storage-aware inference on Apple Silicon.
Hy3 and Qwen MTP backend work, with upstream shipping confirmation and hardware findings in the release discussion.
Read the contribution thread ↗Hy3 and Qwen Mini experiments connect expert pruning and LoRA training to code, tool-format, and memory observations.
Explore the source and receipts ↗iliria explores serving a full GLM expert set from SSD through a C/Metal engine. Its container and the pruned MLX build have different runtime requirements.
Explore the streaming engine ↗Try the work
A code-completion demo and a guide to the wider body of work. Each gives you a practical starting point.
Write the code before and after a cursor, then ask Wisp to fill the gap. A small model trained from scratch to study fill-in-the-middle and native multi-token prediction.
The hosted demo uses the standard 100.7M trunk on CPU. It does not use MTP acceleration or execute generated code.
Find a model, inspect a contribution, load a data sample, or follow a technical guide. The model catalog includes runtime requirements, download sizes, and exact source revisions.
A static guide with no model download required. The source and catalog are available for inspection.
Reusable research assets
Training and calibration material from the GLM Demolition project, with a small MLX-specific sample you can inspect directly.
Code training examples, specialist “soul” material, and calibration corpora used in the pruning and LoRA-healing work.
The original release card reports 272,549 examples across the full JSONL repository. The default viewer indexes a subset; its row total is not the full corpus size.
The card documents thinner Go/JS coverage and a neutral-to-degrading security-soul result on the pruned base. Reuse requires your own evaluation.
A direct connection to MLX
heal/gold2/code_mlx.jsonl contains seven chat-format examples. This snippet downloads only that 58 KB file at the inspected revision and parses it as data.
import json
from pathlib import Path
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="philipjohnbasile/glm52-demolition-data",
repo_type="dataset",
filename="heal/gold2/code_mlx.jsonl",
revision="9c2d273ca6651ca92242cb5ae46bedf0dd8001d0",
)
rows = [json.loads(line) for line in
Path(path).read_text().splitlines()
if line.strip()]
print(len(rows)) # 7
The sample is training material, not a benchmark or a guarantee of code quality. The dataset retains its published license and provenance.
Read, reproduce, build
Existing project documentation, linked at inspected revisions. Follow the implementation details and the evidence behind each release.
Pinned dependencies, restoration steps, and the runtime needed by the Hy3 releases.
The feasibility study behind the smaller-memory Hy3 work. Read the measured limits before choosing a model.
The engine guide for the C/Metal implementation and its SSD-backed model format.
Corpus sources, realized token counts, filtering, and the limits of the surviving provenance record.
The registered code-completion comparison, including null and adverse results and the limits of each protocol.
How the pruning, healing, evaluation, and release pieces fit together in the public source.
The collection / 20 public artifacts
Twenty public artifacts. Clear differences between a model you can try, an experiment, and a component for a larger system.
Download size is not RAM. Totals include root tensor files and sidecars. Runtime, context cache, tokenizer files, and nested adapter checkpoints are separate. How to read the evidence ↓
No models match those filters. Try a broader family or download size.
A note on evidence
These releases include wins, null results, and regressions. The model cards preserve that record.
A head in a checkpoint or a recognized architecture does not prove successful generation. The labels separate a live demo, recorded runs, pending verification, and reference artifacts.
Prompt, token budget, decoding mode, hardware, and harness change the result. Acceptance rate alone is not code quality or speed. Different evaluation protocols are not one leaderboard.
Each catalog entry links to the inspected revision. Model authors, conversion sources, and runtime contributors retain their credits and their own license terms.
Continue exploring
Model weights, data, and evidence are published in versioned repositories. Start with a collection, inspect the files, or join a project discussion.