Teortaxes Questions Legit 80% DeepSeek SWE Scores
AI commentator compares recent model scores to previous overstated reports.
Teortaxes, a pseudonymous DeepSeek booster, notes that early reports of 80% on DeepSWE for ox-alpha later settled at 63%. V4-Pro-0813 officially reaches 62.7% against 12.8% for the Preview version. He states nothing has hit a legitimate 80% yet, though the base model should in theory get there. In reply he adds that Opus 5 and actual DeepSeek models can hit 80% on noisy small samples. He wants the strongest Chinese model release soon, expecting it by late September.
Combined views
3.4K
2 posts, first seen 4h ago
Teortaxes Questions Legit 80% DeepSeek SWE Scores
AI commentator compares recent model scores to previous overstated reports.
Teortaxes, a pseudonymous DeepSeek booster, notes that early reports of 80% on DeepSWE for ox-alpha later settled at 63%. V4-Pro-0813 officially reaches 62.7% against 12.8% for the Preview version. He states nothing has hit a legitimate 80% yet, though the base model should in theory get there. In reply he adds that Opus 5 and actual DeepSeek models can hit 80% on noisy small samples. He wants the strongest Chinese model release soon, expecting it by late September.