DeepSeek V3.2-Speciale Achieves Gold-Medal Results
Narrative
V3.2-Speciale variant delivers gold-medal performance across elite competitions: IMO 2025 (35/42 points), CMO, ICPC World Finals (10/12 problems solved, 2nd place), IOI 2025 (492/600 points). High scores on AIME 2025 (96.0%), HMMT (99.2%). Rivals or exceeds GPT-5-High and Gemini-3.0-Pro on math/coding olympiads.
Reality
Announced with V3.2 launch December 1, 2025. Results independently verifiable via competition archives. Speciale requires more tokens but excels on complex, long-horizon tasks. Temporary API access spurred researcher evaluation; positioned as proof-of-concept for open reasoning at closed-source levels.
Implication
Showcased open-weights models competing at highest academic competition levels. Challenged assumptions on proprietary training data/compute for olympiad mastery. Intensified global debate on AI progress transparency and accessibility. Set benchmark for future agent/math-focused releases.