The MLX Community in 2026 — Where the Real Code Lives

Star counts are vanity

The MLX repo on GitHub has a healthy star count. That tells you almost nothing about whether the framework is actually being maintained, what the release rhythm looks like, or where the useful code is. This lesson is the practical map of the MLX community as of 2026-05 — the repos that matter, the upload zone you should trust, and the rhythm you should track to avoid stale knowledge.

The repos that actually matter

Core framework repos (under the ml-explore org on GitHub) — these are where the framework itself is built:

ml-explore/mlx — the C++/Metal core, Python bindings, the main framework. This is what you pip install mlx.
ml-explore/mlx-examples — reference implementations of common architectures (Llama, Mistral, Qwen, Whisper, Stable Diffusion, MNIST, LoRA, etc.). Read this repo for canonical patterns; many of the ideas in this quest were calibrated against it.
ml-explore/mlx-lm — the language-model serving stack. pip install mlx-lm gives you generation, sampling, the OpenAI-compatible server, and the LoRA / fuse / convert tools. Track 2 of this quest is the practical tour.
ml-explore/mlx-vlm — vision-language models. Smaller surface than mlx-lm but growing.
ml-explore/mlx-audio — TTS and STT. Track 5 lessons 1-2.
ml-explore/mlx-swift — the Swift bindings for shipping inside Apple-platform apps. Track 7 lesson 3.

Where to pull models from

The mlx-community organization on Hugging Face is the trusted upload zone for MLX-format models. When you see a model called mlx-community/Llama-3-8B-Instruct-4bit, that's an MLX-converted, MLX-quantized version of the original Hugging Face model, ready to load with mlx_lm.load. Track 3 covers what to look for in a trustworthy upload (download counts, conversion script, who quantized it).

Random Hugging Face uploaders also publish MLX-format models. Treat those the same way you'd treat any random Docker image — look at the conversion script, look at the upload date, look at whether they linked the source model. The mlx-community org is curated; everywhere else is wild.

The release rhythm to track

MLX ships a release roughly every two weeks. mlx-lm ships somewhat slower, on its own cadence. Both repos use semantic versioning, and the version compatibility between core mlx and mlx-lm is documented in mlx-lm's release notes. The two-week cadence means that something that didn't work three months ago might just work now — and something that worked perfectly last year might have moved to a new API name. The version stamp in this quest's meta.json tells you what I verified against.

What I read to stay current (and what I don't)

Read: GitHub release notes for ml-explore/mlx and ml-explore/mlx-lm. They're concise and accurate.
Read: the issues on those repos when you hit a confusing error — search before you ask.
Read: mlx-examples when you're starting a new project — the canonical pattern is usually there.
Skim: the X / Twitter MLX hashtag for new model uploads. High signal, also high noise.
Don't trust without verifying: random tutorials, blog posts, and especially LLM-generated MLX code. The 2024-era content in particular is full of API names that have since moved. Track 6 lesson 5 has the full survival guide.

Code

Quick versions probe (so you know what you're on)·bash

pip show mlx mlx-lm 2>/dev/null | grep -E "^(Name|Version|Home-page|Summary):"

# Sample (verified 2026-05-03):
#   Name: mlx
#   Version: 0.31.2
#   Name: mlx-lm
#   Version: 0.31.3

Browse the canonical reference repo (clone is optional)·bash

# Browse online: https://github.com/ml-explore/mlx-examples
# If you want a local clone for grep-able reference (skip if you don't):
#   git clone https://github.com/ml-explore/mlx-examples.git ~/code/mlx-examples
# Useful subdirectories to read:
#   llms/         — the canonical LLM serving patterns
#   lora/         — fine-tuning reference
#   stable_diffusion/ — image generation
#   whisper/      — STT reference
#   mnist/        — minimum end-to-end training loop

Exercise

Bookmark these four URLs in your browser, in this order: ml-explore on GitHub, mlx-community on Hugging Face, mlx release notes, mlx-lm release notes. Pick one model from the mlx-community org that you'd like to load in Track 2 (mlx-lm) — write down the exact repo name (mlx-community/...). Pick one model architecture from mlx-examples (e.g. llms, whisper, stable_diffusion) that you find interesting and skim its README.md. You'll come back to both in later tracks.