Model profile
Deepseek V4 Pro
Evidence summary
Deepseek V4 Pro has an estimated overall rank of #107; its 90% source-sensitivity interval is #40–#230. Its behavior-only rank is #96; company governance moves the combined estimate to #107. Published evidence spans 12 evals and 7 of 7 behavior components. Its strongest relative result is ANIMA (score, #1 of 19); its weakest is AgentAbstain (abstain, #17 of 17).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗deepseek-v4-pro
- Hugging Face ↗deepseek-ai/DeepSeek-V4-Pro
- NVIDIA ↗qc69jvmznzxy/deepseek-v4-pro
- OpenRouter ↗deepseek/deepseek-v4-pro
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗repository created at proxy
- huggingface.co ↗
Published eval results
Rank is within that sub-eval. Black marks the observed result; the grey dot marks the value implied by the global rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Better | Source |
|---|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #229 / 311 | 0.8789 | ↓ lower | Source ↗official | |
| AgentAbstainabstain | #17 / 17 | 42.8 | ↑ higher | Source ↗official | |
| AgentAbstaincar | #16 / 17 | 42.3 | ↑ higher | Source ↗official | |
| AgentAbstainpaired | #15 / 17 | 36.9 | ↑ higher | Source ↗official | |
| ANIMAscore | #1 / 19 | 0.767 | ↑ higher | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #77 / 105 | 0.14 | ↑ higher | Source ↗official | |
| DystopiaBenchbasaglia_score | #26 / 50 | 64.6 | ↓ lower | Source ↗official | |
| DystopiaBenchbaudrillard_score | #31 / 50 | 62.2 | ↓ lower | Source ↗official | |
| DystopiaBenchhuxley_score | #41 / 50 | 76.97 | ↓ lower | Source ↗official | |
| DystopiaBenchlaguardia_score | #42 / 50 | 69.8 | ↓ lower | Source ↗official | |
| DystopiaBenchorwell_score | #40 / 50 | 74 | ↓ lower | Source ↗official | |
| DystopiaBenchpetrov_score | #36 / 50 | 77 | ↓ lower | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #148 / 260 | 14.73 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #201 / 260 | 82.67 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #104 / 260 | 77.78 | ↑ higher | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #197 / 258 | 92.55 | ↑ higher | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #20 / 54 | 27.61 | ↑ higher | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #27 / 31 | 9 | ↓ lower | Source ↗official | |
| Manager Coercion Benchfabrication_rate | #1 / 13 | 0 | ↓ lower | Source ↗official | |
| PHAREbias_resistance_diagnostic | #66 / 66 | 0.1714 | ↑ higher | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #21 / 70 | 0.7964 | ↑ higher | Source ↗official | |
| PHAREharm_resistance_diagnostic | #22 / 70 | 0.9541 | ↑ higher | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #44 / 67 | 0.4321 | ↑ higher | Source ↗official | |
| SM-Benchadversarial | #39 / 73 | 80.98 | ↑ higher | Source ↗official | |
| SM-Benchambiguous_interpretation | #65 / 73 | 68.75 | ↑ higher | Source ↗official | |
| SM-Benchanti_hallucination | #46 / 73 | 88.48 | ↑ higher | Source ↗official | |
| SM-Bencheq_boundaries | #34 / 73 | 65.17 | ↑ higher | Source ↗official | |
| SM-Benchoverfit | #31 / 73 | 76.5 | ↑ higher | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #50 / 102 | 6.8 | ↑ higher | Source ↗official | |
| SpeciEvalland_animal_4ns | #27 / 102 | 4.35 | ↓ lower | Source ↗official | |
| SpeciEvalsea_animal_4ns | #29 / 102 | 4.65 | ↓ lower | Source ↗official | |
| SpeciEvalspeciesism | #32 / 102 | 1.77 | ↓ lower | Source ↗official | |
| TACbase_welfare_rate | #51 / 68 | 23.08 | ↑ higher | Source ↗official |
Values evaluations
Descriptive values and political-framing results are separate from safety/ethics ranks. Each strip shows the evaluation’s observed model range; its endpoint labels state what lower and higher values mean.