Big reasoning speedups are here. 1,000× faster training and 280× less compute than any looped model. All done using a fully trainable energy from the 80s
Keep reading to hear how energy is the right paradigm for reasoning
With @HughBlayney@jcrwhittington@prefrontal_ai
🧵/10
@Beamsters1@ddalcu Amazing work! Currently running the mixed-4-8bit model on mlx-serve (which is an absolute joy to see working away), but I can't get more than ~100K tokens. Can't wait to try this on my 96GB M2.
@ddalcu I was experimenting with it to extend session context for Qwen next on 96GB M2 Max. I’m aware of the controversy with original authors of the paper, but was under the impression it was still useful.