MLX-LM has extreme alpha. It’s a unified inference and training framework that works across MLX and CUDA and constantly gets new features without losing modularity. It will only get better.

Ivan Fioravanti ᯅ and now it's free to use omg x.com/YashasGunderia…

always feels weird seeing decentralize baseline outperforming the centralized baseline? i can vaguely remember something similar on diloco when you don't fully decay the learning rate (not 100% sure about this), any thought on why this is happening Teknium (e/λ) emozilla Bowen Peng

Can't run mistralai/Ministral-3-8B-Instruct-2512 locally with MLX, skill issue? 🤷🏻‍♂️ ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported.

to commemorate alan dye moving from apple to meta, here's one of his best quotes

The new Codex model is available in Cursor! It's free to use until December 11th. We worked with OpenAI to optimize Cursor's agent harness for the model. cursor.com/blog/codex-mod…

basketball AI YouTube tutorial is finally live over 1000 hours of work compressed into 37 minutes link below; like and comments please

Agno power!

Why is it so hard to interact in a positive and constructive way on X? We are all here to share and learn from each other. Personally, I’ve learned so many amazing things and discovered so many amazing people here! 🔥

Great GPT-5.1-Codex-Max Prompting Guide!

This is a must have for any MLX lover! And following Gökdeniz Gülmez is another one 🚀

“I am inevitable”

Vibe Infrastructuring

Claude is down! Luckily I have ChatGPT or Grok or Gemini to ask to.

“I am increasingly convinced that the important question is not only what these models can do, but how we learn to work with them.” Thanks for sharing your experience in detail! 🙏

z-image-turbo powered by MLX! 🚀

This is the real superpower of strong founders: being able to skip bureaucracy and fight for power when needed.

mlx-lm can be used for much more than standard LLM training/inference. Here, I use mlx-lm to train an autoregressive 8M transformer on MNIST to generate digits, and further refine it via GRPO w/ self-classification. I use mlx-lm's BatchedGenerate backend for both the GRPO step