Mistral Large 4 on AlphaXiv

Mistral Large 4's authors report that memory-saving attention choices kept language-modeling loss roughly comparable but lowered SWE-bench Verified scores, so they restored long-context attention and then reintroduced sparsity via DeepSeek Sparse Attention. The paper's evidence comes from controlled-scale ablations, and the agentic gain is not attributed to the sparse stage.