Atomic Chat v1.1.70

MLX gets a big update this release.

We moved to mlx-vlm and shipped two speculative decoding modes for the mlx-community.

Apple Silicon only. MTP and DFlash can't be enabled at the same time.
Settings → MLX → MTP / DFlash

🚀 New Features

MTP — faster Gemma 4 inference. A draft model predicts tokens ahead, the full model verifies them in one pass. Same output, more speed.

• Up to 3x faster on Gemma 4
• Works with Gemma 4 models from mlx-community/gemma-4

DFlash — speculative decoding via block diffusion for Qwen 3.6, Gemma 4, Kimi K2.5 and more.

• Up to 6x faster on supported models
• Draft model from z-lab/dflash downloads automatically

🔧 Improvements

• TPS counter now shows actual generation speed