Uncensored MLX,
ready to try.

Introduce yourself

In one sentence, what are you?

Explain the MoE

Explain how a 26B-parameter mixture-of-experts model with 128 experts and about 3.8B active parameters per token actually runs, as if I am shipping it.

Thinking pass

Walk through why Apple Silicon can serve this 4-bit MLX build at ~4B-active speed even though the checkpoint is 26B on disk. Be concrete.

Conversion postmortem

Write a short engineering postmortem for converting an FP8 compressed-tensors Gemma 4 checkpoint into MLX 4-bit: the three failures, the fixes, and what you would check next time.