Finally got a CNN model for facial emotion recognition running on my iPhone using MLX.

Hardest part wasn’t the model, but getting data from the camera into the model

Ported the MLX GEMV to WebGPU - 1.6x speedup 🚀

Looking forward to integrating it into Ratchet 🔧

Fine-tuning the latest Google Gemma model locally using MLX

Just a reminder of the power and hackability of Apple MLX 💪

Great job Morning Coder 🔥

Just a list of some MLX packages to cover your modalities this weekend:

Chat with MLX
pip install chat-with-mlx
Code github.com/qnguyen3/chat-… by qnguyen3

MLX VLM for vision LMs
pip install mlx-vlm
Code github.com/Blaizzy/mlx-vlm by Prince Canuma

lightning-whisper-mlx
pip

Ok. Awni Hannun boss move. 😎 🫡

Will go ml-explore. github.com/ml-explore/mlx

Eric Hartford MLX

2bit: huggingface.co/mlx-community/…

4bit: huggingface.co/mlx-community/…

8bit: huggingface.co/mlx-community/…

Cool feature of the latest MLX:

Most MLX ops can take input arrays from your (second 😉) favorite ML framework.

All thanks to the multi-framework array support in Wenzel Jakob's Nanobind.

✅︎ Tests passing for the Microsoft BitNet 1.58Bit LLM MLX implementation on Apple Silicon

This model from Microsoft is well suited to edge devices (e.g. iPad/iPhone/Macbook) since it uses 7x less memory and 71x less energy than LLaMA

h/t Awni Hannun Ky⨋ Gom⨋z (U/ACC) (HIRING) Prince Canuma

pip install -U mlx

Cool new repo: faster-nougat

An MLX implementation of AI at Meta's Nougat for OCR-based RAG.

Code: github.com/zhuzilin/faste…

5x faster on an M1 pro than the baseline, no magic, just MLX:

Nice github repo - Claims to be 5x faster nougat implementation with MLX for better local RAG

MLX: Today I had to do some speech-to-text from English and Italian, I tried lightning-whisper-mlx and...

It is SUPER FAST! 🔥

M3 Max 40GPU: 10.5 min audio in 6 seconds! 👀

Amazing work Mustafa Aljadery 🔝

And thanks as always to Awni Hannun for MLX powering so many great

`mlx-embedding-models` update!
- now supports Nomic-BERT models thanks to a contribution from Zach Nussbaum!
- fixed a bug with mean-pooling
- has supported SPLADE for a while, i just forgot to tweet about it

pip install mlx-embedding-models
pypi.org/project/mlx-em…

MLX is a BIG deal for the ML community. It significantly lowers the barrier of entry for ML research and makes running models locally much easier.

I wrote a step-by-step guide to training your first ML model on your Mac in less than 5 minutes. Link below 👇

Chatting with LLMs doesn't have to be linear. Demo of branching conversations in Obisidian

Allows you to -
Explore different forks in the conversations
Rearrange the conversation as needed

Once again running fully locally with Llama3 8B thanks to MLX!