deepseek-ai/DeepSeek-V3.2 · Hugging Face (benchmark chart assets/benchmark.png)

Read https://huggingface.co/deepseek-ai/DeepSeek-V3.2 for the benchmark figures on models/deepseek-v3-2. As recorded by the reader: All three values are for the released DeepSeek-V3.2 in thinking mode, read from official DeepSeek sources (not from third-party leaderboard blocks on the card). AIME 2025 93.1 and SWE-bench Verified 73.1 are printed on the official HF model-card benchmark chart; AIME 2025 93.1 and GPQA Diamond 82.4 are printed in the official release-page benchmark table. DeepSeek's published Terminal-Bench row is version 2.0 (46.4), which is not the allowed 'Terminal-Bench 2.1', and its LiveCodeBench row carries no version label (83.3), so neither is reported; MMLU-Pro is not published in these official tables (the card's MMLU-Pro 85 / SWE-bench Verified 70 entries are third-party leaderboard submissions, not lab-published). DeepSeek-V3.2-Speciale values (e.g. AIME 2025 96.0) are a different model and were not used.