Tag
An author improved MLX performance on an M3 Ultra Studio to achieve 73 tok/s for a 4-bit Qwen3.8-Flash-Next model, which shows intelligence scores matching GPT-5.6 Sol, highlighting the growing potential of local AI.