MiniMax-M2.1 open-sourced with strong coding-agent benchmarks

MiniMax released M2.1 with open-source weights (Modified-MIT) and an API, tuned for coding, tool use, and long-horizon planning; it runs locally (llama.cpp requires --jinja) and supports Unsloth Dynamic 2.0 quantization. It posts 74.0 on SWE-bench Verified, 49.4 on Multi-SWE-bench, and 72.5 on multilingual SWE-bench, matching or beating Claude Sonnet 4.5 on several coding tasks and showing stable results across agent frameworks.