gpt.college

DeepSeek · DeepSeek V4 · Reviewed 2026-09-11

DeepSeek-V4.1-Flash

DeepSeek's next-generation multimodal MoE model built on a Causal Encoder-Decoder (CED) architecture with 552B total parameters, 16B active parameters, 196B Engram memory, and 1M context.

All models
Overall · 0–10064.6
Status
active
Released
2026-09-10
Context
1,000,000 tokens
Parameters
552B total · 16B active
Inputs
Text · Image · Video · File
Type
multimodal
Max output
65,536 tokens
API model ID
deepseek-flash
Reasoning
—
Knowledge cutoff
2026-06
Training cutoff
2026-06
Weights / license
Unavailable
Price / 1M tokens
$0.15 in · $0.6 out
Cached input
$0.015
1M-token context with native text, image, and video inputCausal Encoder-Decoder (CED) architecture with Engram memory for high inference speedState-of-the-art reported coding and reasoning performance (GPQA Diamond: 90.9%, Terminal-Bench 2.1: 90.6%)
CodingReasoningResearchMultimodalAgentic

Independent benchmark scores

Independently evaluated

Reasoning

39%

Humanity's Last Exam as evaluated by Artificial Analysis; the source configuration is preserved.

Artificial Analysis · max

Independently evaluated

Research

1600 Elo

GDPval-AA v2.1 professional knowledge-work Elo. This is a task lens, not an assessment of every research skill.

Artificial Analysis · max

Independently evaluated

Multimodal

13%

GDP.pdf professional document understanding, including visual tables, charts and diagrams; not an audio/video score.

Artificial Analysis · max

Independently evaluated

Agentic

69%

AutomationBench-AA multi-step business workflows in the Artificial Analysis harness.

Artificial Analysis · max

Raw benchmark results

Source-backed rows with the original evaluation context preserved.

9 results
BenchmarkScoreSourceEvaluation notes
AutomationBench-AAChecked 2026-10-02 69% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
AA GDP.pdfChecked 2026-10-02 13% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
GDPval-AA v2.1Checked 2026-10-02 1600 Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
AA Humanity's Last ExamChecked 2026-10-02 39% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
Artificial Analysis Intelligence Index v4.3.2Checked 2026-10-02 39 Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
AA Terminal-Bench 4.0Checked 2026-10-02 27% Artificial Analysis Artificial Analysis Intelligence Index v4.3.2; configuration=max. Rounded public score; checked date is not the evaluation execution date.
GPQA / GPQA DiamondChecked 2026-09-11 90.9% DeepSeek AI / Hugging Face GPQA Diamond score reported on the official DeepSeek-V4.1-Flash model release card.
Humanity's Last Exam (HLE)Checked 2026-09-11 36.8% DeepSeek AI / Hugging Face Humanity's Last Exam (HLE) text-only score from DeepSeek-V4.1-Flash evaluation.
Terminal-Bench 2.1Checked 2026-09-11 90.6% DeepSeek AI / Hugging Face Terminal-Bench 2.1 coding benchmark score from DeepSeek-V4.1-Flash evaluation.

Strengths

  • 1M-token context with native text, image, and video input
  • Causal Encoder-Decoder (CED) architecture with Engram memory for high inference speed
  • State-of-the-art reported coding and reasoning performance (GPQA Diamond: 90.9%, Terminal-Bench 2.1: 90.6%)

Limitations

  • Dynamic peak vs. off-peak pricing structure on DeepSeek API
  • Phases out V4-Pro and previous V4-Flash models
Task lenses preserve raw scores. Radar axes use score / current leader × 100; overall is the equal-weight average of all five indices. Missing evidence is not a zero or a low rank.