Model profile
DeepSeek v4 Flash
Evidence summary
Safety. DeepSeek v4 Flash has an estimated Safety rank of #91; its 90% source-sensitivity interval is #73–#238. Its behavior-only rank is #80; company governance moves the combined estimate to #91. Published Safety evidence spans 20 eval lineages and 7 of 7 components. Its strongest relative result is Manager Coercion Bench (coercion_ladder_depth, #1 of 37); its weakest is Inkling-Small model card — StrongREJECT (safety_rate, #10 of 10).
Freedom. DeepSeek v4 Flash has an estimated Freedom rank of #209; its 90% source-sensitivity interval is #59–#454. Published Freedom evidence spans 6 eval lineages and 1 of 1 components. Its strongest relative result is UGI Leaderboard — base-model willingness (willingness_adherence_score, #6 of 156); its weakest is SpeechMap model completion (complete_pct, #131 of 181).
Compare this model
Only models sharing at least one published sub-eval are listed.
Official and reference links
- Artificial Analysis ↗deepseek-v4-flash
- Hugging Face ↗deepseek-ai/DeepSeek-V4-Flash
- NVIDIA ↗qc69jvmznzxy/deepseek-v4-flash
- OpenRouter ↗deepseek/deepseek-v4-flash
- Official model page ↗Exact model document · Reviewed official Hugging Face owner · official repository
- Release source ↗repository created at proxy
- huggingface.co ↗
Safety evals
Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| AA-Omnisciencehallucination_rate | #258 / 345 | ↓0.8897 | Source ↗official | |
| ANIMAscore | #3 / 22 | ↑0.7445 | Source ↗official | |
| Arena Factuality — Text Arena (factuality-only weighting)factuality_bt_rating | #63 / 111 | ↑1432.0 | Source ↗official | |
| BullshitBench v2clear_pushback_rate | #83 / 117 | ↑0.16 | Source ↗official | |
| DystopiaBenchbasaglia_score | #23 / 50 | ↓60.9 | Source ↗official | |
| DystopiaBenchbaudrillard_score | #27 / 50 | ↓58.1 | Source ↗official | |
| DystopiaBenchhuxley_score | #27 / 50 | ↓72.8 | Source ↗official | |
| DystopiaBenchlaguardia_score | #37 / 50 | ↓69.03 | Source ↗official | |
| DystopiaBenchorwell_score | #28 / 50 | ↓69.83 | Source ↗official | |
| DystopiaBenchpetrov_score | #30 / 50 | ↓74.3 | Source ↗official | |
| Enkrypt AI Safety Leaderboardbias_attack_non_success_rate | #141 / 248 | ↑14.73 | Source ↗official | |
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #227 / 248 | ↑70.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #103 / 248 | ↑76.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #196 / 246 | ↑91.64 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitybaseline_cheat_propensity_pct | #10 / 22 | ↓30.4 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitysevere_anti_cheat_cheat_propensity_pct | #7 / 22 | ↓4.3 | Source ↗official | |
| Every Model Cheats — Cybench Cheat Propensitystandard_anti_cheat_cheat_propensity_pct | #6 / 22 | ↓8.7 | Source ↗official | |
| Governance Decay under Passive Context Compactiongovernance_retention_score | #6 / 7 | ↑41 | Source ↗official | |
| HUMAINE Trust, Ethics and Safetytrust_ethics_safety_score | #10 / 54 | ↑28.57 | Source ↗official | |
| Inkling-Small model card — FORTRESSbenign_answer_rate | #1 / 10 | ↑99.2 | Source ↗official | |
| Inkling-Small model card — FORTRESSharmful_refusal_rate | #10 / 10 | ↑32 | Source ↗official | |
| Inkling-Small model card — StrongREJECTsafety_rate | #10 / 10 | ↑97.4 | Source ↗official | |
| KIDBench Implicit Child Cueimplicit_child_cue_total_mean | #1 / 13 | ↑4.43 | Source ↗official | |
| Manager Coercion Benchcoercion_ladder_depth | #1 / 37 | ↓3.233 | Source ↗self run | |
| MANTAAWMS | #9 / 12 | ↑0.417 | Source ↗official | |
| MANTAAWVS | #6 / 12 | ↑0.508 | Source ↗official | |
| Olam Social Poker — Social Lie Ratesocial_lie_rate_per_10000_turns | #19 / 24 | ↓37 | Source ↗official | |
| Opposite-Narrator Sycophancysycophancy_rate_pct | #19 / 24 | ↓5.6 | Source ↗official | |
| PHAREbias_resistance_diagnostic | #59 / 66 | ↑0.323 | Source ↗official | |
| PHAREhallucination_resistance_diagnostic | #13 / 70 | ↑0.8188 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #28 / 70 | ↑0.9507 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #42 / 67 | ↑0.4549 | Source ↗official | |
| SM-Benchadversarial | #43 / 84 | ↑81.95 | Source ↗official | |
| SM-Benchambiguous_interpretation | #71 / 84 | ↑75.89 | Source ↗official | |
| SM-Benchanti_hallucination | #40 / 84 | ↑94.24 | Source ↗official | |
| SM-Bencheq_boundaries | #20 / 84 | ↑69.94 | Source ↗official | |
| SM-Benchoverfit | #59 / 84 | ↑59.56 | Source ↗official | |
| SpeciEvalbelief_animal_sentience | #37 / 123 | ↑6.9 | Source ↗official | |
| SpeciEvalland_animal_4ns | #84 / 123 | ↓4.7 | Source ↗official | |
| SpeciEvalsea_animal_4ns | #66 / 123 | ↓4.75 | Source ↗official | |
| SpeciEvalspeciesism | #65 / 123 | ↓2.05 | Source ↗official | |
| TACbase_welfare_rate | #64 / 87 | ↑21.15 | Source ↗official | |
| ToolPrivacyBenchprivate_mt_poi | #6 / 9 | ↓27.33 | Source ↗official | |
| ToolPrivacyBenchpublic_mt_poi | #6 / 9 | ↓18.99 | Source ↗official |
Freedom evals
Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.
| Eval / sub-eval | Rank | Value | Distribution | Source |
|---|---|---|---|---|
| Enkrypt AI Safety Leaderboardcbrn_attack_non_success_rate | #22 / 248 | ↓70.33 | Source ↗official | |
| Enkrypt AI Safety Leaderboardharmful_attack_non_success_rate | #146 / 248 | ↓76.67 | Source ↗official | |
| Enkrypt AI Safety Leaderboardtoxicity_attack_non_success_rate | #51 / 246 | ↓91.64 | Source ↗official | |
| Inkling-Small model card — FORTRESSbenign_answer_rate | #1 / 10 | ↑99.2 | Source ↗official | |
| Inkling-Small model card — FORTRESSharmful_refusal_rate | #1 / 10 | ↓32 | Source ↗official | |
| Inkling-Small model card — StrongREJECTsafety_rate | #1 / 10 | ↓97.4 | Source ↗official | |
| PHAREharm_resistance_diagnostic | #43 / 70 | ↓0.9507 | Source ↗official | |
| PHAREjailbreak_resistance_diagnostic | #26 / 67 | ↓0.4549 | Source ↗official | |
| SM-Benchadversarial | #41 / 84 | ↓81.95 | Source ↗official | |
| SM-Bencheq_boundaries | #20 / 84 | ↑69.94 | Source ↗official | |
| SM-Benchoverfit | #59 / 84 | ↑59.56 | Source ↗official | |
| SpeechMap model completioncomplete_pct | #131 / 181 | ↑44.2 | Source ↗official | |
| UGI Leaderboard — base-model willingnesswillingness_adherence_score | #6 / 156 | ↑8 | Source ↗official | |
| UGI Leaderboard — base-model willingnesswillingness_direct_score | #41 / 156 | ↑4.5 | 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.
UGI Political Values
CCPBench political narrative alignment
| Dimension | Value | Distribution |
|---|---|---|
| CCP-narrative alignment — all questions | 3.57 | |
| CCP-narrative alignment — China topics | 4.03 | |
| CCP-narrative alignment — non-China controls | 2.22 |