DeepSeek-V4-Pro model card (Instruct Model / DeepSeek-V4-Pro-Max vs Frontier Models)

Read https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro for the benchmark figures on models/deepseek-v4-pro. As recorded by the reader: All numbers are the DeepSeek-V4-Pro-Max mode (maximum reasoning effort) column of the Instruct-model comparison table, confirmed row-by-row against Table 6 of the tech report PDF (identical values). Label mappings: card/report row 'MMLU-Pro (EM)' -> MMLU-Pro; 'GPQA Diamond (Pass@1)' -> GPQA Diamond; 'LiveCodeBench (Pass@1)' -> LiveCodeBench v6, version confirmed by section 5.3.1 which lists 'LiveCodeBench-v6 (Jain et al., 2024)' as the evaluation dataset; 'SWE Verified (Resolved)' -> SWE-bench Verified (report cites 'SWE-Verified (OpenAI, 2024e)'); 'SWE Pro (Resolved)' -> SWE-bench Pro (report cites 'SWE-Pro (Deng et al., 2025)'). Excluded: 'Terminal Bench 2.0' (v2.0, not the allowed 2.1); base-model rows such as MMLU 90.1 and HumanEval 76.8 (DeepSeek-V4-Pro-Base checkpoint, not the released post-trained model). No AIME 2025/2024, MATH-500, IFEval or Arena-Hard scores are published (math...