DeepSeek V3.1 Terminus: benchmark results
DeepSeek's V3.1 Terminus update with improved agentic tool use and language consistency. Provider: DeepSeek. Released 2025-09-22. Access: Open.
Aggregate Rating 1690 ± 8, rank #220 of 1656 rated models, from 344 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FireBench - Overconfidence | 86 | Share of the 370 overconfidence samples (%) in which the mod | 100 |
| SWE-bench Live Java (62 tasks, SWE-agent) | 27.42 | Resolved rate (%) | 100 |
| UGI Leaderboard | 57.58 | UGI Score | 97.8 |
| AGC-Bench - ss_gen | 1.34 | Dataset z-score | 97.5 |
| TuRTLe - Verilator Syntax | 95.25 | Average Score (%) | 95.5 |
| JP-TL-Bench - English to Japanese - Easy | 9.96 | LT Score (0-10; Easy items only; anchored pairwise Bradley-T | 95.2 |
| AGC-Bench - conceptual_design | 1.3 | Dataset z-score | 93.9 |
| JP-TL-Bench - English to Japanese | 9.94 | LT Score (0-10; anchored pairwise Bradley-Terry vs 20 Base S | 93.7 |
| JP-TL-Bench - English to Japanese - Hard | 9.94 | LT Score (0-10; Hard items only; anchored pairwise Bradley-T | 93.7 |
| TuRTLe - Icarus Syntax | 93.7 | Average Score (%) | 93.2 |
| LiveFMBench - 2025 | 13.64 | Pass@1 (%) after the faithfulness check on LiveFMBench-2025 | 92.9 |
| AGC-Bench - tinystories | 1.23 | Dataset z-score | 92.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-1-terminus · How It Works · Data refreshed daily, snapshot 2026-10-20.