RLTasksby ulam.ai
Models / DeepSeek

DeepSeek / MODEL EVIDENCE

DeepSeek V4.1 Flash (none)

4 benchmarks · 4 observations · 4 sources · released 2026-09-09

Scores at a glance

4 benchmarks · large dot: DeepSeek V4.1 Flash (none); small dots: closest models where reported

CritPt ↗0.29%
032.29

DeepSeek V4.1 Flash (none) · CritPt · Accuracy · % ↗

SciCode ↗35.53%
1.2766.9

DeepSeek V4.1 Flash (none) · SciCode · Score · % ↗

Closest models by score

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

6.4 percentile-point gap

4 of 4 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 →

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. SciCode Benchmark | Artificial Analysis ↗

    Artificial Analysis

  3. ARC-AGI-2 ↗

    Epoch AI

  4. CritPt v1 ↗

    Epoch AI

Names used in source tables

DeepSeek V4.1 Flash (none)