Episode Details

Ditch the "Frankenstein" architecture for native multimodality. We're dissecting the MiniMax M3, from its "Step Zero" training and million-token context window to its brutal dominance over GPT-5.5 and Gemini 3.1 Pro on the SWE-Bench. Is the M3's ability to autonomously optimize CUDA kernels a productivity miracle or a death knell for high-end engineering? Join us as we dive into Sparse Attention, Mixture-of-Experts, and the shifting power dynamics of the global AI arms race.