Legal Research Bench ↗23.08 %
DeepSeek V4 · Legal Research Bench · All-Pass ↗
DeepSeek / MODEL EVIDENCE
DeepSeek V4
22 benchmarks · 23 observations · 23 sources
Scores at a glance
22 benchmarks · large dot: DeepSeek V4; small dots: closest models where reported
LegalBench ↗80.32 %
DeepSeek V4 · LegalBench · Accuracy ↗
DeepSeek V4 · Harvey's Legal Agent Benchmark v1 · Accuracy ↗
CaseLaw v2 ↗59.38 %
DeepSeek V4 · CaseLaw · Accuracy ↗
Tax Agent Bench ↗25.17 %
DeepSeek V4 · Tax Agent Bench · Accuracy ↗
Vals Finance Agent v2 ↗44.08 %
DeepSeek V4 · Finance Agent v2 · Accuracy ↗
DeepSeek V4 · Excel Modeling Benchmark v1 · Accuracy ↗
CorpFin v2 ↗61.38 %
DeepSeek V4 · CorpFin · Accuracy ↗
MortgageTax & TaxEval v2 ↗72.08 %
DeepSeek V4 · TaxEval · Accuracy ↗
MedCode ↗40.45 %
DeepSeek V4 · MedCode · Accuracy ↗
MedScribe ↗75.14 %
DeepSeek V4 · MedScribe · Accuracy ↗
ProofBench ↗16 %
DeepSeek V4 · ProofBench · Accuracy ↗
GPQA Diamond ↗89.39 %
DeepSeek V4 · GPQA Diamond · Accuracy ↗
MMLU-Pro ↗87.25 %
DeepSeek V4 · MMLU Pro · Accuracy ↗
Vibe Code Bench ↗49.93 %
DeepSeek V4 (OpenHands) · Vibe Code Bench · original task set ↗
DeepSeek V4 · ProgramBench · Accuracy ↗
Terminal-Bench 4.0 ↗11.11 %
DeepSeek V4 (Mini-SWE-agent) · Terminal-Bench 4.0 · Score ↗
SkillsBench ↗51.27 %
DeepSeek V4 (OpenHands) · SkillsBench · Score ↗
Code Migration v1.1 ↗26.2 %
DeepSeek V4 · Code Migration v1 · Accuracy ↗
SWE-bench Verified ↗77.4 %
DeepSeek V4 (Mini-SWE-agent) · SWE-bench Verified · Score ↗
LiveCodeBench ↗87.48 %
DeepSeek V4 · LiveCodeBench · Accuracy ↗
Public Benefits Bench v1.1 ↗62.92 %
DeepSeek V4 · Public Benefits Bench · Accuracy ↗
1 more reported result
57.58 % · DeepSeek V4 · Public Benefits Bench v1 · Accuracy · Vals AI · Public Benefits Bench v1 ↗
Closest models by score
Similar positions on shared benchmark leaderboards. Smaller gaps are closer.
DeepSeek V4 Flash 0731 →DeepSeek
16 of 22 shared benchmark comparisons
Kimi K2.6 →Kimi
17 of 22 shared benchmark comparisons
DeepSeek V4 Pro 0813 →DeepSeek
16 of 22 shared benchmark comparisons
How similarity is measured
Mean absolute gap in percentile positions on shared benchmark settings; lower is closer. At least 60% overlap is required, with a minimum of three when available. Independent evaluations are used when there are at least three; otherwise published report comparisons are used. Repeated model observations within one setting are averaged before ranking. Writing-style traits and composite indices are excluded. This describes similarity, not an overall quality ranking.
Sources
Every plotted observation links to its evaluation. Reported configurations, benchmark versions and task subsets can differ; similarity describes shared scores, not training lift.
- CaseLaw — evaluation methodology and results ↗
- Code Migration v1 — evaluation methodology and results ↗
- CorpFin — evaluation methodology and results ↗
- Excel Modeling Benchmark v1 — evaluation methodology and results ↗
- Finance Agent v2 — evaluation methodology and results ↗
- GPQA Diamond — evaluation methodology and results ↗
- Harvey's Legal Agent Benchmark v1 — evaluation methodology and results ↗
- LiveCodeBench — evaluation methodology and results ↗
- LegalBench — evaluation methodology and results ↗
- Legal Research Bench — evaluation methodology and results ↗
- MedCode — evaluation methodology and results ↗
- MedScribe — evaluation methodology and results ↗
- MMLU Pro — evaluation methodology and results ↗
- ProgramBench — evaluation methodology and results ↗
- ProofBench — evaluation methodology and results ↗
- Public Benefits Bench — evaluation methodology and results ↗
- Public Benefits Bench v1 — evaluation methodology and results ↗
- SkillsBench — evaluation methodology and results ↗
- SWE-bench Verified — evaluation methodology and results ↗
- Tax Agent Bench — evaluation methodology and results ↗
- TaxEval — evaluation methodology and results ↗
- Terminal-Bench 4.0 — evaluation methodology and results ↗
- Vibe Code Bench — evaluation methodology and results ↗
Names used in source tables
DeepSeek V4 · DeepSeek V4 (Mini-SWE-agent) · DeepSeek V4 (OpenHands)