[whatmodel.sh](../../README.md) / DeepSeek V4.1 Flash # DeepSeek V4.1 Flash DeepSeek · released 2026-09 · 552B params (16B active) · mit **score 75.5** ±0.6 · #29 of 732 · from 34 benchmark results across 5 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at ● marks DeepSeek V4.1 Flash among every model on that task or test, worst to best. 
task                vs the others         rank  counts  tests
agent tasks         ▁▁▃▆█▄▅▅▄●▂▁    #20 of 313     26%     10
coding              ▁▁▁▃▅▇█▇▅▄●▁    #26 of 469    8.0%      2
chat & writing      ▁▁▁▃▆▄▆█▆▇●▃   #28 of 402*    3.6%      1
math                ▁▁▃▅█▆▄█●▂▁▁    #36 of 431     32%      4
vision & documents  ▁▁▂▂▁▃▄▂▆▇●▄   #42 of 153*    2.4%      1
web search          ██·········●      #1 of 3*    1.2%      1
agent tasks · #20 of 313 · 26% of its score Beats [Muse Wonder](muse-wonder.md), [GPT-5.5](gpt-5-5.md), [Hy4 preview (770B-A49B)](hy4-preview.md) here, though they rank higher overall. 
test                                    result  vs the others in its run  difficulty            counts  why
Arena Agent (October 2026 aggrega…  0.0191 IPS  ▁▁▁▃▃▂●▁▂▂▁▁ beats 35/59  avg -0.379 IPS          9.1%  sharp
Terminal-Bench 4.0 (maker)               31.2%  ▁·▁··▁▆▆●██▃ beats 16/33  very hard · avg 0.7%    5.8%  sharp · at its level · small run · seve…
Chartography with tools (maker)          78.9%  ▂········●██ beats 2/10   very hard · avg 9.2%    2.5%  sharp · small run
Terminal-Bench 2.1 (maker)               90.6%  ▁·▁·▁▁▄▃▆█▅● beats 62/63  hard · avg 29%          2.1%  sharp · small run
DeepSWE v1.1 (resolved) (maker)          74.2%  ▂····▂▁▁▆██● beats 28/31  hard · avg 23%          1.7%  at its level · small run · several runs
Video2World V2WScore (coding agen…       30.5%  █··▄●·▄····▄ beats 3/6    very hard · avg 7.3%    1.5%  at its level
Humanity's Last Exam with tools (…       63.9%  ▁·▁▂▂▁▄▄▇█▄● beats 56/62  hard · avg 16%          1.2%  at its level · small run
Agents' Last Exam (pass@1, offici…       31.8%  █··········● beats 2/2    very hard · avg 7.0%    0.8%  at its level · small run
Agents' Last Exam (metric/harness…       31.8%  ▂·▂·█●▂·▂▂▄▄ beats 7/14   very hard · avg 7.9%    0.7%  at its level · small run
Terminal-Bench 3.0 (maker)               30.0%  █··█·······● beats 2/2    very hard · avg 2.5%    0.5%  noisy · small run
coding · #26 of 469 · 8.0% of its score Beats [GPT-5.6 Sol](gpt-5-6-sol.md), [GPT-5.5](gpt-5-5.md), [GPT-5.6 Terra](gpt-5-6-terra.md) here, though they rank higher overall. Trails [Grok 4.7](grok-4-7.md), [Qwen3.8-Flash-Next](qwen3-8-flash-next.md), [Qwen3.8 Max (0902)](qwen3-8-max-0902.md) here, though they rank lower overall. 
test                             result  vs the others in its run    difficulty      counts  why
Arena WebDev (Code Arena)        1619.2  ▁▁▃▂▅█▃▆●▅▁▂ beats 104/129  avg 981           6.3%  sharp
DeepSWE v1.1 (resolved) (maker)   74.2%  ▂····▂▁▁▆██● beats 28/31    hard · avg 23%    1.7%  at its level · small run · several runs
chat & writing · #28 of 402 · 3.6% of its score Beats [GPT-6 Astra](gpt-6-astra.md), [GPT-6 Sol](gpt-6-sol.md), [GPT-5.6 Terra](gpt-5-6-terra.md) here, though they rank higher overall. Trails [Gemini 3.6 Flash](gemini-3-6-flash.md), [GLM-5.3-Flash](glm-5-3-flash.md), [Gemini 3.5 Flash](gemini-3-5-flash.md) here, though they rank lower overall. 
test            result  vs the others in its run    difficulty  counts  why
LMArena (text)  1476.2  ▁▁▁▂▃▃▃▆▆▇●▃ beats 356/385  avg 1260      3.6%  blunt
math · #36 of 431 · 32% of its score Trails [GPT-5.2](gpt-5-2.md), [Qwen3.7 Max](qwen3-7-max.md), [Gemini 3.1 Pro (Preview)](gemini-3-1-pro.md) here, though they rank lower overall. 
test                             result  vs the others in its run    difficulty            counts  why
FrontierMath tier 4 v2            26.8%  ▁▂▁▂●▅▂▃▂▁▁▁ beats 28/67    very hard · avg 0.1%     16%  sharp · at its level
FrontierMath tiers 1–3 v2         67.4%  ▂··▁▁▃▄▄█●▄▃ beats 55/86    very hard · avg 1.2%     14%  sharp · at its level
MathArena Apex (pass@1) (maker)   65.6%  ▄····▄···▄●▄ beats 4/5      hard · avg 17%          1.3%  at its level · small run
OTIS mock AIME 2024–2025          98.3%  ▃▃▄▃▃▃▇▇█▃●▅ beats 174/197  very hard · avg 7.2%    0.4%  sharp · maxed out · noisy
vision & documents · #42 of 153 · 2.4% of its score Beats [MiMo-V2.6-Pro](mimo-v2-6-pro.md), [MiMo-V2.6-Flash (309B-A15B)](mimo-v2-6-flash.md) here, though they rank higher overall. Trails [Qwen3.6 Flash](qwen3-6-flash.md), [Gemini 3.5 Flash-Lite](gemini-3-5-flash-lite.md), [Qwen3.6 Plus](qwen3-6-plus.md) here, though they rank lower overall. 
test                             result  vs the others in its run    difficulty  counts  why
Arena vision (style controlled)  1261.4  ▁▁▂▂▂▂▃▄▅▅●▇ beats 114/152  avg 1061      2.4%  blunt
web search · #1 of 3 · 1.2% of its score 
test                                result  vs the others in its run  difficulty      counts  why
Humanity's Last Exam with tools (…   63.9%  ▁·▁▂▂▁▄▄▇█▄● beats 56/62  hard · avg 16%    1.2%  at its level · small run
other tests (13) 
test                                result  vs the others in its run    difficulty            counts  why
Mystery Game Puzzles (Epoch)         43.0%  █··▂▃▇▄█●▄▂▁ beats 62/77    very hard · avg 9.4%    6.3%  sharp · at its level · noisy
Furniture Assembly (Epoch)           34.2%  █▆●▆▆▂▆▂▂▂▂▅ beats 11/32    very hard · avg 1.2%    6.3%  sharp · at its level
GPQA Diamond                         89.8%  ▁▁▁▁▂▄▆▄▅▇●▅ beats 173/211  hard · avg 38%          5.5%  sharp · small run · several runs
Humanity's Last Exam (maker)         39.1%  ▁▁▁▄▆▆▇██●▁▃ beats 94/111   very hard · avg 5.4%    3.3%  at its level · small run · several runs
NL2Repo (maker)                      64.0%  ▂▂▂·▃▃▅█▃▂▅● beats 21/22    very hard · avg 15%     2.5%  at its level · small run · several runs
AutomationBench v1.0.6 (maker)       54.8%  ▃···▃·▃▃··█● beats 7/8      very hard · avg 9.9%    1.2%  at its level · small run
SEC-Bench Pro (pass@1) (maker)       62.8%  ▄·▄·▄·●▄···▄ beats 4/6      very hard · avg 8.7%    1.0%  at its level · small run
CyberGym (maker)                     88.1%  ▄···▄··▄█▄█● beats 8/8      medium · avg 40%        0.9%  small run
BabyVision with tools (maker)        89.6%  █······█··●█ beats 2/3      medium · avg 49%        0.8%  small run
ExploitGym (pass@1) (maker)          15.3%  ▄·▄··█·●▄··▄ beats 4/6      very hard · avg 0.7%    0.8%  noisy · small run
AutomationBench (DeepSeek V4.1 ca…   54.8%  █··█·······● beats 2/2      very hard · avg 13%     0.8%  at its level · small run
ProgramBench (Almost@1) (maker)      20.3%  █●·······█·█ beats 1/3      very hard · avg 5.2%    0.6%  small run
ZeroBench main with tools (pass@5…   49.0%  ▄········●·▄ beats 1/3                                 —  not counted yet: no other model in the …
## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its *counts* column above; here it is added up by who ran the tests. [Method](../method.md) 
source                                    results  share of the score
Epoch AI                                        6   45% ▇▇▇▇▇
LMArena leaderboard dataset                     4   21% ▇▇▇
deepseek-ai/DeepSeek-V4.1-Flash model c…       15   20% ▇▇
DeepSeek V4.1 Flash 1017 model card (su…        8   12% ▇
Video2World: Benchmarking Coding Agents…        1    1% ▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 75.5
Licences: Epoch AI ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · deepseek-ai/DeepSeek-V4.1-Flash model card ([MIT](../../licenses/deepseek-ai--DeepSeek-V4.1-Flash.txt)) · DeepSeek V4.1 Flash 1017 model card ([MIT](../../licenses/deepseek-1017-card.txt)) · Video2World: Benchmarking Coding Agents for Interactive World Modeling from Embodied Videos ([CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/))
Not counted (3) 
source                  test                            score  why
deepseek-ai/DeepSeek-…  BabyVision with tools           89.6%  only model in this source's run
deepseek-ai/DeepSeek-…  Chartography with tools         78.9%  only model in this source's run
deepseek-ai/DeepSeek-…  ZeroBench main with tools (pa…  49.0%  too few models have it yet
## How to run it ### Other ways to run it 1. Run it as published (fp8 + fp4 experts, needs ~510.3 GB, NVIDIA only, from [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/2cba9e42aa026125f3ed06c6d98c1db82f7ca027/inference/README.md)) ```sh hf download deepseek-ai/DeepSeek-V4.1-Flash --local-dir DeepSeek-V4.1-Flash-HF && cd DeepSeek-V4.1-Flash-HF/inference && python -m pip install -r requirements.txt && python convert.py --hf-ckpt-path .. --save-path ../../DeepSeek-V4.1-Flash-TP8 --model-parallel 8 --expert-dtype fp4 --tokenizer-path .. && torchrun --nproc-per-node 8 generate.py --ckpt-path ../../DeepSeek-V4.1-Flash-TP8 --config config.json --interactive --temperature 0.6 ``` Model repository and documentation: [huggingface.co/deepseek-ai](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) ## Published speed measurements Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine. 
hardware                    precision               speed  engine and settings
2× NVIDIA H200, 768 GB RAM  GSQ-RCO 3bit-CP…  12.08 tok/s  llama.cpp 5210c7c5ed61, CUDA · ll…
Each speed links to its source, which has the full settings. [interactive version](https://whatmodel.sh/model/deepseek-v4-1-flash) · [JSON](https://whatmodel.sh/api?format=json&all=1)