DeepSeek V3: benchmark results
Uses compressed attention state to reduce inference memory and predicts multiple tokens during training. Its post-training distills reasoning from an R1-series teacher into shorter responses. Provider: DeepSeek. Released 2024-12-26. Access: Open.
Aggregate Rating 1644 ± 4, rank #321 of 1656 rated models, from 1383 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AIM-Bench (inventory manager) | 14.72 | BG Avg. C (self-reported) | 100 |
| BBH | 87.5 | Score (imported) | 100 |
| BesiegeField - Catapult - Single-agent | 3.5 | Mean distance (m) | 100 |
| BioDesignBench (Guided Mode) | 58.46 | Overall Score | 100 |
| DITING | 5.16 | Score | 100 |
| DITING - Lexical Ambiguity | 5.52 | Score | 100 |
| DITING - Tense Consistency | 5.46 | Score | 100 |
| DITING - Terminology Localization | 5.12 | Score | 100 |
| Diagnosing LLM Arbitration Behavior over Pre-e | 23.5 | Margin (self-reported) | 100 |
| EComStage - Attitude Classification | 88.68 | Task Score (%) | 100 |
| EComStage - Query Rewrite | 82.49 | Task Score (%) | 100 |
| FinBen - FNS | 37.72 | Normalized Score | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3 · How It Works · Data refreshed daily, snapshot 2026-10-20.