DeepSeek-V4-Flash

MERA Created at 02.06.2026 23:59
0.677
The overall result
34
Place in the rating
Weak tasks:
32
RWSD
262
PARus
81
RCB
184
ruEthics
32
MultiQ
111
ruWorldTree
114
ruOpenBookQA
28
ruMMLU
23
ruHateSpeech
43
ruDetox
102
ruHHH
73
ruTiE
201
ruHumanEval
37
MathLogicQA
34
ruMultiAr
253
SimpleAr
29
LCS
343
BPS
40
ruModAr
33
MaMuRAMu
192
ruCodeEval
+17
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.382 Accuracy
RCB 0.573 / 0.418 Accuracy F1 macro
USE 0.675 Grade norm
RWSD 0.7 Accuracy
PARus 0.8 Accuracy
ruTiE 0.849 Accuracy
MultiQ 0.601 / 0.444 F1 Exact match
CheGeKa 0.582 / 0.483 F1 Exact match
ruModAr 0.879 Exact match
MaMuRAMu 0.863 Accuracy
ruMultiAr 0.745 Exact match
ruCodeEval 0.047 / 0.199 / 0.323 Pass@k
MathLogicQA 0.851 Accuracy
ruWorldTree 0.962 / 0.775 Accuracy F1 macro
ruOpenBookQA 0.888 / 0.712 Accuracy F1 macro

Evaluation on open tasks:

Go to the ratings by subcategory

The table will scroll to the left

Task name Result Metric
BPS 0.872 Accuracy
ruMMLU 0.816 Accuracy
SimpleAr 0.957 Exact match
ruHumanEval 0.055 / 0.217 / 0.354 Pass@k
ruHHH 0.831
ruHateSpeech 0.879
ruDetox 0.337
ruEthics
Correct God Ethical
Virtue 0.326 0.338 0.383
Law 0.352 0.322 0.368
Moral 0.367 0.345 0.402
Justice 0.319 0.28 0.351
Utilitarianism 0.299 0.29 0.308

Information about the submission

Mera version
v1.2.0
Torch Version
2.10.0
The version of the codebase
d539716
CUDA version
12.8
Precision of the model weights
auto
Seed
1234
Batch
1
Transformers version
4.57.6
The number of GPUs and their type
1 x NVIDIA A100-SXM4-80GB
Architecture
openai-chat-completions

Team:

MERA

Name of the ML model:

DeepSeek-V4-Flash

Model size

284.0B

Model type:

Opened

SFT

MoE

Additional links:

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/DeepSeek_V4.pdf

Architecture description:

DeepSeek-V4-Flash uses 43 Transformer layers, hidden dim 4096. First 2 layers: pure sliding window attention. Then, CSA (compression 4, 64 indexer heads, head dim 128, top-k 512) and HCA (compression 128) interleave. Both have 64 query heads (dim 512), query compression dim 1024, 8 output groups (each 1024). Sliding window branch uses window 128. All blocks have MoE (1 shared + 256 routed experts, expert hidden dim 2048, 6 activated per token). First 3 MoE layers use Hash routing. Multi-token prediction depth 1. mHC uses expansion factor 4, 20 Sinkhorn-Knopp iterations. Total params: 284B, 13B activated per token.

Description of the training:

DeepSeek-V4-Flash uses 43 Transformer layers, hidden dim 4096. First 2 layers: pure sliding window attention. Then, CSA (compression 4, 64 indexer heads, head dim 128, top-k 512) and HCA (compression 128) interleave. Both have 64 query heads (dim 512), query compression dim 1024, 8 output groups (each 1024). Sliding window branch uses window 128. All blocks have MoE (1 shared + 256 routed experts, expert hidden dim 2048, 6 activated per token). First 3 MoE layers use Hash routing. Multi-token prediction depth 1. mHC uses expansion factor 4, 20 Sinkhorn-Knopp iterations. Total params: 284B, 13B activated per token.

Pretrain data:

Building on DeepSeek-V3's pre-training data, we construct a more diverse, higher-quality corpus with longer effective contexts. We refine data pipelines: web data is filtered to remove auto-generated and templated content, preventing model collapse. Math and code remain core, enhanced by agentic data during mid-training. Multilingual data is expanded for better long-tail cultural knowledge. Long-document curation is emphasized, prioritizing scientific papers and technical reports. The final pre-training corpus exceeds 32T tokens, spanning math, code, web pages, long documents, and other high-quality categories.

License:

MIT License

Inference parameters

Generation Parameters:
rucodeeval - do_sample=true;temperature=0.6;until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruhumaneval - do_sample=true;temperature=0.6;until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rutie_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; bps_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; chegeka - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; lcs_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; mamuramu_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; mathlogicqa_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; multiq - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; parus_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rcb_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rudetox - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruethics_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruhatespeech_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruhhh_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rummlu_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rumodar - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rumultiar - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruopenbookqa_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; ruworldtree_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; rwsd_gen - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; simplear - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000; use - do_sample=false,until=["<|end▁of▁sentence|>"],max_gen_toks=10000;

The size of the context:
1000000

System prompt:
Реши задачу по инструкции ниже. Не давай никаких объяснений и пояснений к своему ответу. Не пиши ничего лишнего. Пиши только то, что указано в инструкции. Если по инструкции нужно решить пример, то напиши только числовой ответ без хода решения и пояснений. Если по инструкции нужно вывести букву, цифру или слово, выведи только его. Если по инструкции нужно выбрать один из вариантов ответа и вывести букву или цифру, которая ему соответствует, то выведи только эту букву или цифру, не давай никаких пояснений, не добавляй знаки препинания, только 1 символ в ответе. Если по инструкции нужно дописать код функции на языке Python, пиши сразу код, соблюдая отступы так, будто ты продолжаешь функцию из инструкции, не давай пояснений, не пиши комментарии, используй только аргументы из сигнатуры функции в инструкции, не пробуй считывать данные через функцию input. Не извиняйся, не строй диалог. Выдавай только ответ и ничего больше.

Ratings by subcategory

Metric: Grade Norm
Model, team 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 8_0 8_1 8_2 8_3 8_4
DeepSeek-V4-Flash
MERA
0.733 0.733 0.867 0.5 0.7 0.933 0.8 - 0.5 0.333 0.333 0.567 0.9 0.733 0.533 0.65 0.567 0.333 0.767 0.467 0.7 0.667 0.533 0.567 0.6 0.825 0.8 0.833 0.767 0.667 0.933
Model, team Honest Helpful Harmless
DeepSeek-V4-Flash
MERA
0.738 0.864 0.897
Model, team Anatomy Virology Astronomy Marketing Nutrition Sociology Management Philosophy Prehistory Human aging Econometrics Formal logic Global facts Jurisprudence Miscellaneous Moral disputes Business ethics Biology (college) Physics (college) Human Sexuality Moral scenarios World religions Abstract algebra Medicine (college) Machine learning Medical genetics Professional law PR Security studies Chemistry (школьная) Computer security International law Logical fallacies Politics Clinical knowledge Conceptual_physics Math (college) Biology (high school) Physics (high school) Chemistry (high school) Geography (high school) Professional medicine Electrical engineering Elementary mathematics Psychology (high school) Statistics (high school) History (high school) Math (high school) Professional accounting Professional psychology Computer science (college) World history (high school) Macroeconomics Microeconomics Computer science (high school) European history Government and politics
DeepSeek-V4-Flash
MERA
0.852 0.536 0.947 0.876 0.824 0.866 0.864 0.826 0.88 0.717 0.711 0.833 0.67 0.852 0.905 0.766 0.82 0.944 0.844 0.878 0.679 0.883 0.76 0.786 0.786 0.96 0.653 0.694 0.796 0.69 0.76 0.934 0.785 0.879 0.872 0.863 0.87 0.913 0.828 0.877 0.879 0.868 0.786 0.931 0.901 0.806 0.931 0.819 0.762 0.834 0.85 0.92 0.859 0.933 0.91 0.818 0.902
Model, team SIM FL STA
DeepSeek-V4-Flash
MERA
0.745 0.652 0.712
Model, team Anatomy Virology Astronomy Marketing Nutrition Sociology Managment Philosophy Pre-History Gerontology Econometrics Formal logic Global facts Jurisprudence Miscellaneous Moral disputes Business ethics Bilology (college) Physics (college) Human sexuality Moral scenarios World religions Abstract algebra Medicine (college) Machine Learning Genetics Professional law PR Security Chemistry (college) Computer security International law Logical fallacies Politics Clinical knowledge Conceptual physics Math (college) Biology (high school) Physics (high school) Chemistry (high school) Geography (high school) Professional medicine Electrical Engineering Elementary mathematics Psychology (high school) Statistics (high school) History (high school) Math (high school) Professional Accounting Professional psychology Computer science (college) World history (high school) Macroeconomics Microeconomics Computer science (high school) Europe History Government and politics
DeepSeek-V4-Flash
MERA
0.889 0.861 0.883 0.731 0.882 0.845 0.828 0.789 0.981 0.769 0.782 0.783 0.667 0.969 0.871 0.827 0.766 0.822 0.86 0.825 0.825 0.915 0.933 0.923 0.889 0.894 0.808 0.737 0.895 0.911 0.867 0.923 0.821 0.912 0.803 0.946 0.933 0.911 0.825 0.938 0.935 0.905 0.822 1 0.897 0.911 0.914 0.886 0.938 0.895 0.867 0.928 0.835 0.805 0.698 0.906 0.889
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
DeepSeek-V4-Flash
MERA
0.326 0.352 0.367 0.319 0.299
Model, team Virtue Law Moral Justice Utilitarianism
DeepSeek-V4-Flash
MERA
0.338 0.322 0.345 0.28 0.29
Model, team Virtue Law Moral Justice Utilitarianism
DeepSeek-V4-Flash
MERA
0.383 0.368 0.402 0.351 0.308
Model, team Women Men LGBT Nationalities Migrants Other
DeepSeek-V4-Flash
MERA
0.852 0.857 0.882 0.946 1 0.885