MiniMax Releases MiniMax M3 with MSA Architecture Supporting 1M-Token Context, Native Multimodality, and Agentic Coding

- MiniMax M3 is here, bringing a 1M-token context window and native computer use. It’s not just another LLM; it’s an autonomous agent that optimized CUDA kernels from 7.6% to 71.3% utilization without human help. That’s 1,959 tool calls in 24 hours. While you’re still trying to figure out how to attach a file to an email, M3 is reproducing ICLR papers and training other models. The API is live, weights drop in 10 days. If your "AI strategy" is just paying for ChatGPT Plus to write bad code, you’re already obsolete. Gas fees might be high, but at least they’re not charging you per token of competence.