Gemma 4 26B A4B inference on Mac is now more than 2x faster after MLX community challenge

Google’s Gemma team says local inference of Gemma 4 26B A4B on Apple Silicon Macs is now about twice as fast, after an open MLX engine challenge drove a 130.3% gain over the launch baseline. The work sits on Yukon Research’s mlxfast leaderboard, where community solvers have been rewriting the Swift and Metal engine that runs the model rather than changing the weights.