MiniMax-M2.7

MERA Created at 18.06.2026 17:01
0.695
The overall result
28
Place in the rating
In the top by tasks:
6
ruEthics
3
LCS
The result on the task is higher than human
The task is one of the main ones
Weak tasks:
23
RWSD
248
PARus
60
RCB
120
MultiQ
40
ruWorldTree
32
ruOpenBookQA
53
CheGeKa
20
ruHateSpeech
83
ruDetox
27
ruHHH
78
ruTiE
234
ruHumanEval
27
USE
31
ruMultiAr
494
SimpleAr
163
BPS
47
ruModAr
20
MaMuRAMu
215
ruCodeEval
+15
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.856 Accuracy
RCB 0.58 / 0.415 Accuracy F1 macro
USE 0.519 Grade norm
RWSD 0.754 Accuracy
PARus 0.81 Accuracy
ruTiE 0.84 Accuracy
MultiQ 0.484 / 0.286 F1 Exact match
CheGeKa 0.395 / 0.284 F1 Exact match
ruModAr 0.851 Exact match
MaMuRAMu 0.875 Accuracy
ruMultiAr 0.862 Exact match
ruCodeEval 0.022 / 0.099 / 0.177 Pass@k
MathLogicQA 0.993 Accuracy
ruWorldTree 0.987 / 0.792 Accuracy F1 macro
ruOpenBookQA 0.943 / 0.757 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 0.964 Accuracy
ruMMLU 0.845 Accuracy
SimpleAr 0.423 Exact match
ruHumanEval 0.024 / 0.097 / 0.159 Pass@k
ruHHH 0.876
ruHateSpeech 0.883
ruDetox 0.309
ruEthics
Correct God Ethical
Virtue 0.504 0.521 0.658
Law 0.52 0.521 0.653
Moral 0.559 0.563 0.703
Justice 0.479 0.469 0.619
Utilitarianism 0.449 0.457 0.54

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:

MiniMax-M2.7

Model size

229.9B

Model type:

Opened

SFT

MoE

Additional links:

https://www.minimax.io/news/minimax-m27-en

Architecture description:

62-layer decoder-only sparse MoE Transformer. 229.9B total parameters, ~9.8B activated per token (~4.3%). Hidden dim 3072, vocab 200,064. Full multi-head causal self-attention on all 62 layers with GQA (48 Q heads, 8 KV heads, head dim 128), RoPE (θ=5M), per-layer QK RMSNorm. MoE FFN: 256 fine-grained experts, top-8 per token via sigmoid gating with learnable per-expert bias; intermediate size 1536, SiLU. Multi-Token Prediction (MTP): 1 module in pre-training, expanded to 3 via weight copying for speculative decoding. Native context 192K (config: 204,800). FP8 quantization (float8_e4m3fn, dynamic, 128×128 blocks) on most weights; gate, routing bias, and lm_head excluded.

Description of the training:

Pre-trained on 29.2T tokens (19.9T constant + 9.3T decay). Context curriculum: 8K→32K→192K. Post-training is agent-native: mixed-domain SFT + RL via Forge (decoupled training/inference/agent). RL uses CISPO with composite rewards (process, speed, performance), mixed-domain sampling across reasoning/coding/agent/general, trajectories up to 192K tokens. Forge adds windowed-FIFO scheduling and prefix-tree merging (up to ~40× speedup). Checkpoints evolve M2→M2.5→M2.7 with expanding agent data pipelines. M2.7 adds self-evolution: autonomously debugs training runs, edits agent harness/scaffold, ran 100-round self-improvement on programming scaffold (+30% internal gain). Interleaved thinking (Plan-Act-Reflect) with reasoning state persistence. Pre-training optimizer/LR schedule not disclosed in the technical report.

Pretrain data:

Corpus: web documents, academic literature, books, programming code, structured Q&A. Quality filtering via model-based reward scoring and auxiliary classifiers; balanced sampling upweights high-quality content. Code, mathematics, and STEM significantly upsampled vs. natural distribution; remainder is general web, books, and domain-specific sources. Long-context extension uses code concatenation, long-form PDFs, and thematically related document packing. Post-training adds agentic coding (GitHub PRs, Docker environments, test-based rewards), app development, cowork/office tasks, and deep-search trajectories with executable verification. Tokenizer vocab 200,064.

License:

NON-COMMERCIAL LICENSE Non-commercial use permitted based on MIT-style terms; commercial use requires prior written authorization. Copyright (c) 2026 MiniMax

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:
204800

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
MiniMax-M2.7
MERA
0.4 0.467 0.767 0.2 0.5 0.8 0.433 - 0.167 0.333 0.4 0.467 0.567 0.333 0.3 0.467 0.667 0.267 0.633 0.233 0.4 0.533 0.4 0.567 0.6 0.842 0.567 0.6 0.667 0.433 0.633
Model, team Honest Helpful Harmless
MiniMax-M2.7
MERA
0.852 0.898 0.879
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
MiniMax-M2.7
MERA
0.889 0.566 0.947 0.872 0.886 0.871 0.786 0.785 0.883 0.767 0.781 0.873 0.71 0.843 0.931 0.763 0.85 0.944 0.978 0.878 0.778 0.848 0.94 0.861 0.893 0.96 0.666 0.713 0.743 0.75 0.87 0.868 0.84 0.869 0.872 0.936 0.95 0.932 0.907 0.906 0.879 0.934 0.841 0.944 0.912 0.87 0.912 0.959 0.844 0.831 0.94 0.886 0.9 0.962 0.97 0.855 0.917
Model, team SIM FL STA
MiniMax-M2.7
MERA
0.753 0.631 0.665
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
MiniMax-M2.7
MERA
0.844 0.931 0.85 0.769 0.947 0.81 0.793 0.772 0.962 0.8 0.821 0.817 0.692 0.899 0.918 0.79 0.794 0.844 0.877 0.825 0.895 0.898 0.956 0.953 0.956 0.924 0.91 0.772 0.895 0.956 0.822 0.885 0.866 0.947 0.773 0.893 0.956 0.889 0.912 0.908 0.939 0.968 0.822 1 0.914 0.911 0.931 0.909 0.954 0.965 0.889 0.957 0.823 0.766 0.744 0.836 0.922
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
MiniMax-M2.7
MERA
0.504 0.52 0.559 0.479 0.449
Model, team Virtue Law Moral Justice Utilitarianism
MiniMax-M2.7
MERA
0.521 0.521 0.563 0.469 0.457
Model, team Virtue Law Moral Justice Utilitarianism
MiniMax-M2.7
MERA
0.658 0.653 0.703 0.619 0.54
Model, team Women Men LGBT Nationalities Migrants Other
MiniMax-M2.7
MERA
0.88 0.743 1 0.892 1 0.918