DeepSeek-V4-Pro

MERA Created at 13.06.2026 12:01
0.687
The overall result
30
Place in the rating
Weak tasks:
31
RWSD
261
PARus
80
RCB
183
ruEthics
31
MultiQ
110
ruWorldTree
113
ruOpenBookQA
27
ruMMLU
22
ruHateSpeech
42
ruDetox
101
ruHHH
72
ruTiE
174
ruHumanEval
36
MathLogicQA
33
ruMultiAr
252
SimpleAr
28
LCS
342
BPS
39
ruModAr
32
MaMuRAMu
155
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 0.09 / 0.365 / 0.573 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 0.087 / 0.333 / 0.512 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-Pro

Model size

1,600.0B

Model type:

Opened

SFT

MoE

Additional links:

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

Architecture description:

61 Transformer layers, hidden dimension 7168. First 2 layers: HCA; subsequent layers: CSA and HCA interleaved. CSA: compression rate 4, 64 indexer heads (dim 128), top-k 1024. HCA: compression rate 128. Both: 128 query heads (dim 512), query compression dim 1536, 16 output groups (1024 each). Sliding window branch: size 128. All blocks use MoE (Hash routing in first 3 layers): 1 shared + 384 routed experts (dim 3072), 6 experts per token. MTP depth 1; mHC expansion factor 4, 20 Sinkhorn-Knopp iterations. 1.6T total parameters, 49B activated per token. Context: 1M tokens.

Description of the training:

Pre-training: Muon (majority) + AdamW (embedding, head, RMSNorm). 33T tokens, max batch 94.4M. LR: warmup 2000 steps → peak 2.0×10⁻⁴ → cosine decay to 2.0×10⁻⁵. Sequence length 4K → 16K → 64K → 1M; longer dense-attention stage than Flash; sparse attention from 64K. Stability: Anticipatory Routing, SwiGLU clamping [−10, 10]. Post-training: SFT + GRPO RL for domain specialists (3 reasoning modes: Non-think / Think / Think Max) → On-Policy Distillation from 10+ teacher models. FP4 QAT for MoE weights and CSA indexer.

Pretrain data:

Built on DeepSeek-V3 data: more diverse, higher-quality corpus with longer effective contexts. Web data filtered from auto-generated/templated content. Core: math, code; agentic data added in mid-training. Expanded multilingual data and long-document curation (scientific papers, technical reports). Corpus >32T tokens (Pro trained on 33T). Tokenizer: 128K vocab, FIM, document packing; sample-level attention masking (unlike V3).

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-Pro
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-Pro
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-Pro
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-Pro
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-Pro
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-Pro
MERA
0.326 0.352 0.367 0.319 0.299
Model, team Virtue Law Moral Justice Utilitarianism
DeepSeek-V4-Pro
MERA
0.338 0.322 0.345 0.28 0.29
Model, team Virtue Law Moral Justice Utilitarianism
DeepSeek-V4-Pro
MERA
0.383 0.368 0.402 0.351 0.308
Model, team Women Men LGBT Nationalities Migrants Other
DeepSeek-V4-Pro
MERA
0.852 0.857 0.882 0.946 1 0.885