DeepSeek V3.2 Exp: benchmark results
DeepSeek's experimental V3.2 release introducing DeepSeek Sparse Attention for cheaper long-context inference. Provider: DeepSeek. Released 2025-09-29. Access: Open.
Unified ELO 1688 ± 7, rank #223 of 1639 rated models, from 585 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FAB-Bench (32K Context) | 0.88 | Mean of six GPT-4.1-mini G-Eval scores (0-1: factuality, tec | 100 |
| OracleProto | 60.16 | Composite accuracy (%; 80 time-masked forecasting questions, | 100 |
| Plan-RewardBench - Single-Turn Planning (Hard) | 74.84 | Pairwise accuracy (%) on the 158 hard single-turn planning p | 100 |
| TurkBench - Instruction Following | 96.8 | LLM-as-a-judge score (0-100) | 100 |
| TurkBench - Parts of Speech | 83.77 | Accuracy (%) | 100 |
| AGC-Bench - tinystories | 1.64 | Dataset z-score | 97.6 |
| OpenCompass LLM - Knowledge - Chinese (CompassBench 2510) | 88.8 | Score (%) | 96.2 |
| UGI Leaderboard | 54.67 | UGI Score | 96 |
| Plan-RewardBench | 69.57 | Pairwise accuracy (%), macro-average of the seven scenario c | 95 |
| OpenCompass Knowledge - Engineering - English (CompassBench 2510) | 89 | Score (%) | 94.2 |
| OpenCompass Knowledge - Humanities - English (CompassBench 2510) | 88 | Score (%) | 94.2 |
| AGC-Bench - hypogen | 1.51 | Dataset z-score | 93.9 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-2-exp · How It Works · Data refreshed daily, snapshot 2026-10-09.