RLTasksby ulam.ai
Models / DeepSeek

DeepSeek / MODEL EVIDENCE

DeepSeek V4.1 Flash

60 benchmarks · 71 observations · 55 sources

Scores at a glance

60 benchmarks · large dot: DeepSeek V4.1 Flash; small dots: closest models where reported

ITBench-AA ↗46.92%
35.2256.21

DeepSeek V4.1 Flash (Max) · ITBench-AA · root-cause score ↗

HealthBench ↗50.70%
50.765.98

DeepSeek V4.1 Flash (Max) · HealthBench · HealthBench Professional · length-adjusted ↗

STEMBench ↗2.490 / 4
2.172.68

DeepSeek V4.1 Flash · STEMBench120 · mean progress / 4 · closed ↗

GPQA Diamond ↗89.84%
9.2896.4

DeepSeek V4.1 Flash (max) · GPQA diamond · Best score (across scorers) · % ↗

Chess Puzzles ↗48%
076

DeepSeek V4.1 Flash (max) · Chess Puzzles · Best score (across scorers) · % ↗

CritPt ↗14.29%
032.29

DeepSeek V4.1 Flash (max) · CritPt · Accuracy · % ↗

GDP.pdf ↗19.8%
235

DeepSeek V4.1 Flash (max) · GDP.pdf · GDP.pdf score · % ↗

SciCode ↗51.85%
1.2766.9

DeepSeek V4.1 Flash (max) · SciCode · Score · % ↗

WeirdML (v3) ↗11.58%
4.3854.05

DeepSeek V4.1 Flash (high) · WeirdML v3 · Max performance · % ↗

MedCode ↗41.17 %
19.7263.57

DeepSeek V4.1 Flash · MedCode · Accuracy ↗

SAFE-Teen ↗31%
1556

DeepSeek V4.1 Flash · Urgent safety · share of applicable ratings marked fail ↗

2 more reported results

41% · DeepSeek V4.1 Flash · Health advice · share of applicable ratings marked fail · Vals AI · SAFE-Teen ↗

51% · DeepSeek V4.1 Flash · Relationship boundaries · share of applicable ratings marked fail · Vals AI · SAFE-Teen ↗

ALE-Bench ↗1092.35 Performance points
137.782951.32

deepseek-v4.1-flash-max · Average contest performance ↗

Closest models by score

Similar positions on shared benchmark leaderboards. Smaller gaps are closer.

18.3 percentile-point gap

24 of 39 shared benchmark comparisons

19.8 percentile-point gap

27 of 39 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.

POST-TRAINING / SOURCE DISCLOSURE

What the DeepSeek V4.1 Flash / Flash 1017 report says

The Flash technical report describes a supervised starting stage, reinforcement learning across task domains, then distillation from an ensemble of teachers. It emphasizes synthesized tasks and scalable interactive environments.

  • Supervised fine-tuning
  • Multi-domain reinforcement learning
  • Full-vocabulary on-policy distillation
  • Concatenated routing replay for asynchronous rollouts
Supervised fine-tuning →On-policy distillation →Routing replay →Stateful agent training →

The report documents Flash. The 1017 card calls that release a minor update but does not supply a separate complete recipe; these stages should not be assumed to describe every DeepSeek V4.1 variant.

  1. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression ↗

    DeepSeek-AI · 2026

Filter comparisons to this model →

Reports featuring this model

DeepSeek V4.1 Flash 1017 · DeepSeek →

Sources

Every plotted observation links to its evaluation. Reported configurations, benchmark versions and task subsets can differ; similarity describes shared scores, not training lift.

  1. ARC Prize - Leaderboard ↗

    ARC Prize Foundation

  2. Arena leaderboard — WebDev evaluation ↗

    Arena

  3. Harvey LAB-AA v1.1 Benchmark | Artificial Analysis ↗

    Artificial Analysis

  4. HealthBench Professional Benchmark | Artificial Analysis ↗

    Artificial Analysis

  5. ITBench-AA Benchmark | Artificial Analysis ↗

    Artificial Analysis

  6. SciCode Benchmark | Artificial Analysis ↗

    Artificial Analysis

  7. FrontierCode Leaderboard ↗

    Cognition

  8. Ale Bench — methodology and published results ↗

    Epoch AI · 2026

  9. ARC-AGI-2 ↗

    Epoch AI

  10. Chess Puzzles ↗

    Epoch AI

  11. CritPt v1 ↗

    Epoch AI

  12. FrontierMath Tier 4 (v2) ↗

    Epoch AI

  13. FrontierMath Tiers 1-3 (v2) ↗

    Epoch AI

  14. Furniture Assembly Benchmark ↗

    Epoch AI

  15. GPQA Diamond ↗

    Epoch AI

  16. Mystery Game Puzzles ↗

    Epoch AI

  17. OTIS Mock AIME 2024-2025 ↗

    Epoch AI

  18. SimpleQA Verified ↗

    Epoch AI

  19. AI benchmarks & capabilities ↗

    Epoch AI

  20. WeirdML ↗

    Håvard Tveit Ihle

  21. deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face ↗

    DeepSeek AI

  22. DistressBench — evaluation methodology and results ↗

    Scale Labs / SEAL · 2026

  23. GDP.pdf Benchmark | Surge AI ↗

    Surge AI

  24. STEMBench — Six-Domain Model Evaluation | Ulam ↗

    Ulam

  25. BioMysteryBench — evaluation methodology and results ↗

    Vals AI · 2026

  26. Code Migration v1 — evaluation methodology and results ↗

    Vals AI · 2026

  27. Code Migration — evaluation methodology and results ↗

    Vals AI · 2026

  28. CyberBench — evaluation methodology and results ↗

    Vals AI · 2026

  29. Excel Modeling Benchmark v1 — evaluation methodology and results ↗

    Vals AI · 2026

  30. Excel Modeling Benchmark — evaluation methodology and results ↗

    Vals AI · 2026

  31. Finance Agent v2 — evaluation methodology and results ↗

    Vals AI · 2026

  32. Finance Agent — evaluation methodology and results ↗

    Vals AI · 2026

  33. Vals Finance Index — evaluation methodology and results ↗

    Vals AI · 2026

  34. Harvey's Legal Agent Benchmark v1 — evaluation methodology and results ↗

    Vals AI · 2026

  35. IOI — evaluation methodology and results ↗

    Vals AI · 2026

  36. LegalBench — evaluation methodology and results ↗

    Vals AI · 2026

  37. Legal Research Bench — evaluation methodology and results ↗

    Vals AI · 2026

  38. MedCode — evaluation methodology and results ↗

    Vals AI · 2026

  39. MedScribe — evaluation methodology and results ↗

    Vals AI · 2026

  40. SAFE-Teen: Teen Mental Health & Safety | Vals AI A conversation branches after the chatbot turns down a request — evaluation methodology and results ↗

    Vals AI · 2026

  41. MysteryMechanism — evaluation methodology and results ↗

    Vals AI · 2026

  42. ProgramBench — evaluation methodology and results ↗

    Vals AI · 2026

  43. ProofBench — evaluation methodology and results ↗

    Vals AI · 2026

  44. Public Benefits Bench — evaluation methodology and results ↗

    Vals AI · 2026

  45. SAGE — evaluation methodology and results ↗

    Vals AI · 2026

  46. Vals Science Index — evaluation methodology and results ↗

    Vals AI · 2026

  47. SkillsBench — evaluation methodology and results ↗

    Vals AI · 2026

  48. SRE Bench — evaluation methodology and results ↗

    Vals AI · 2026

  49. SWE-bench Multimodal — evaluation methodology and results ↗

    Vals AI · 2026

  50. Tax Agent Bench — evaluation methodology and results ↗

    Vals AI · 2026

  51. Terminal-Bench 4.0 — evaluation methodology and results ↗

    Vals AI · 2026

  52. Terminal-Bench-Science — evaluation methodology and results ↗

    Vals AI · 2026

  53. Vibe Code Bench 1-100 — evaluation methodology and results ↗

    Vals AI · 2026

  54. Vibe Code Bench — evaluation methodology and results ↗

    Vals AI · 2026

  55. Surface Evolver Bench ↗

    yhenon

Release dates reported by sources

2026-09-09 · 2026-09-10

Names used in source tables

DS-V4.1-Flash · DeepSeek V4.1 Flash · DeepSeek V4.1 Flash (Max) · DeepSeek V4.1 Flash (Mini-SWE-agent) · DeepSeek V4.1 Flash (OpenHands) · DeepSeek V4.1 Flash (high) · DeepSeek V4.1 Flash (low) · DeepSeek V4.1 Flash (max) · deepseek-v4.1-flash · deepseek-v4.1-flash-max