MiniMax M3
MiniMax M3 is an open-weight flagship model released June 1, 2026, using MiniMax Sparse Attention to cut compute costs to roughly 1/20 at a 1M-token context, with frontier coding capabilities. Weights are available ungated at MiniMaxAI/MiniMax-M3 (~155K downloads), but note the license is a bespoke minimax-community license, not MIT, so review its terms before commercial use; the SWE-bench Pro score should be disregarded due to benchmark concerns.