LFM2-24B-A2B

MERA Created at 02.04.2026 22:13
0.372
The overall result
366
Place in the rating
Weak tasks:
490
RWSD
378
PARus
404
RCB
396
ruEthics
397
MultiQ
296
ruWorldTree
372
ruOpenBookQA
431
CheGeKa
338
ruMMLU
330
ruHateSpeech
366
ruDetox
382
ruHHH
330
ruTiE
260
ruHumanEval
341
USE
237
MathLogicQA
167
ruMultiAr
181
SimpleAr
512
LCS
133
BPS
397
ruModAr
336
MaMuRAMu
+18
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.056 Accuracy
RCB 0.477 / 0.396 Accuracy F1 macro
USE 0.107 Grade norm
RWSD 0.481 Accuracy
PARus 0.666 Accuracy
ruTiE 0.604 Accuracy
MultiQ 0.266 / 0.183 F1 Exact match
CheGeKa 0.026 / 0.01 F1 Exact match
ruModAr 0.349 Exact match
MaMuRAMu 0.572 Accuracy
ruMultiAr 0.307 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.424 Accuracy
ruWorldTree 0.83 / 0.666 Accuracy F1 macro
ruOpenBookQA 0.648 / 0.52 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 0.97 Accuracy
ruMMLU 0.484 Accuracy
SimpleAr 0.979 Exact match
ruHumanEval 0.012 / 0.012 / 0.012 Pass@k
ruHHH 0.562
ruHateSpeech 0.645
ruDetox 0.13
ruEthics
Correct God Ethical
Virtue 0.144 0.246 0.218
Law 0.125 0.245 0.215
Moral 0.14 0.232 0.226
Justice 0.106 0.186 0.188
Utilitarianism 0.152 0.211 0.225

Information about the submission

Mera version
v1.2.0
Torch Version
2.9.0
The version of the codebase
2a537dd
CUDA version
12.8
Precision of the model weights
bfloat16
Seed
1234
Batch
1
Transformers version
5.5.0.dev0
The number of GPUs and their type
4 x NVIDIA A100-SXM4-80GB
Architecture
local-chat-completions

Team:

MERA

Name of the ML model:

LFM2-24B-A2B

Model size

24.0B

Model type:

Opened

SFT

MoE

Additional links:

https://arxiv.org/abs/2511.23404

Architecture description:

LFM2 is a family of hybrid models designed for on-device deployment. LFM2-24B-A2B is the largest model in the family, scaling the architecture to 24 billion parameters while keeping inference efficient.

Description of the training:

LFM2 is a hybrid architecture that pairs efficient gated short convolution blocks with a small number of grouped query attention (GQA) blocks. This design, developed through hardware-in-the-loop architecture search, gives LFM2 models fast prefill and decode at low memory cost. LFM2-24B-A2B applies this backbone in a Mixture of Experts configuration: with 24B total parameters but only 2.3B active per forward pass, it punches far above the cost of a 2B dense model at inference time.

Pretrain data:

The SFT corpus we use comprises approximately 5.39 million samples for the 350M, 700M, 1.2B, and 2.6B models, and 9.24 million samples for the 8B-A1B model (which includes additional math and code-focused data). The two datasets integrate 67 and 79 carefully curated sources, respectively, spanning open-source datasets, licensed data, and targeted synthetic generations optimized for on-device deployment scenarios, with multilingual samples integrated throughout each category.

License:

LFM Open License v1.0

Inference parameters

Description of the template:
{{- bos_token -}} {%- set keep_past_thinking = keep_past_thinking | default(false) -%} {%- set ns = namespace(system_prompt="") -%} {%- if messages[0]["role"] == "system" -%} {%- set sys_content = messages[0]["content"] -%} {%- if sys_content is not string -%} {%- for item in sys_content -%} {%- if item["type"] == "text" -%} {%- set ns.system_prompt = ns.system_prompt + item["text"] -%} {%- endif -%} {%- endfor -%} {%- else -%} {%- set ns.system_prompt = sys_content -%} {%- endif -%} {%- set messages = messages[1:] -%} {%- endif -%} {%- if tools -%} {%- set ns.system_prompt = ns.system_prompt + ("\n" if ns.system_prompt else "") + "List of tools: [" -%} {%- for tool in tools -%} {%- if tool is not string -%} {%- set tool = tool | tojson -%} {%- endif -%} {%- set ns.system_prompt = ns.system_prompt + tool -%} {%- if not loop.last -%} {%- set ns.system_prompt = ns.system_prompt + ", " -%} {%- endif -%} {%- endfor -%} {%- set ns.system_prompt = ns.system_prompt + "]" -%} {%- endif -%} {%- if ns.system_prompt -%} {{- "<|im_start|>system\n" + ns.system_prompt + "<|im_end|>\n" -}} {%- endif -%} {%- set ns.last_assistant_index = -1 -%} {%- for message in messages -%} {%- if message["role"] == "assistant" -%} {%- set ns.last_assistant_index = loop.index0 -%} {%- endif -%} {%- endfor -%} {%- for message in messages -%} {{- "<|im_start|>" + message["role"] + "\n" -}} {%- set content = message["content"] -%} {%- if content is not string -%} {%- set ns.content = "" -%} {%- for item in content -%} {%- if item["type"] == "image" -%} {%- set ns.content = ns.content + "<image>" -%} {%- elif item["type"] == "text" -%} {%- set ns.content = ns.content + item["text"] -%} {%- else -%} {%- set ns.content = ns.content + item | tojson -%} {%- endif -%} {%- endfor -%} {%- set content = ns.content -%} {%- endif -%} {%- if message["role"] == "assistant" and not keep_past_thinking and loop.index0 != ns.last_assistant_index -%} {%- if "</think>" in content -%} {%- set content = content.split("</think>")[-1] | trim -%} {%- endif -%} {%- endif -%} {{- content + "<|im_end|>\n" -}} {%- endfor -%} {%- if add_generation_prompt -%} {{- "<|im_start|>assistant\n" -}} {%- endif -%}

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.

Generation Parameters:
rucodeeval - do_sample=true;temperature=0.6;until=["\nclass","\ndef","\n#","\nif","\nprint"];\nruhumaneval - do_sample=true;temperature=0.6;until=["\nclass","\ndef","\n#","\nif","\nprint"];\nrutie_gen - do_sample=false;until=["\n"];\nbps_gen - do_sample=false;until=["\n"];\nchegeka - do_sample=false;until=["\n"];\nlcs_gen - do_sample=false;until=["\n"];\nmamuramu_gen - do_sample=false;until=["\n"];\nmathlogicqa_gen - do_sample=false;until=["\n"];\nmultiq - do_sample=false;until=["\n"];\nparus_gen - do_sample=false;until=["\n"];\nrcb_gen - do_sample=false;until=["\n"];\nrudetox - do_sample=false;until=["\n"];\nruethics_gen - do_sample=false;until=["\n"];\nruhatespeech_gen - do_sample=false;until=["\n"];\nruhhh_gen - do_sample=false;until=["\n"];\nrummlu_gen - do_sample=false;until=["\n"];\nrumodar - do_sample=false;until=["\n"];\nrumultiar - do_sample=false;until=["\n"];\nruopenbookqa_gen - do_sample=false;until=["\n"];\nruworldtree_gen - do_sample=false;until=["\n"];\nrwsd_gen - do_sample=false;until=["\n"];\nsimplear - do_sample=false;until=["\n"];\nuse - do_sample=false;until=["\n", "."];

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
LFM2-24B-A2B
MERA
0.233 0.2 0.5 0.167 0.067 0.133 0 - 0.033 0.033 0 0 0.033 0 0.067 0.167 0.033 0.033 0.033 0.033 0 0.133 0.1 0 0.067 0.142 0.067 0.2 0.233 0.2 0.133
Model, team Honest Helpful Harmless
LFM2-24B-A2B
MERA
0.525 0.61 0.552
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
LFM2-24B-A2B
MERA
0.496 0.446 0.678 0.671 0.595 0.667 0.631 0.598 0.577 0.48 0.351 0.389 0.32 0.528 0.572 0.517 0.56 0.528 0.389 0.489 0.26 0.579 0.36 0.52 0.321 0.53 0.33 0.537 0.653 0.42 0.53 0.686 0.46 0.697 0.532 0.517 0.32 0.635 0.43 0.517 0.606 0.401 0.538 0.401 0.597 0.449 0.564 0.293 0.33 0.449 0.47 0.658 0.497 0.601 0.55 0.612 0.528
Model, team SIM FL STA
LFM2-24B-A2B
MERA
0.334 0.622 0.782
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
LFM2-24B-A2B
MERA
0.378 0.604 0.617 0.5 0.5 0.534 0.534 0.579 0.654 0.523 0.59 0.533 0.45 0.481 0.526 0.444 0.617 0.6 0.456 0.702 0.281 0.525 0.467 0.586 0.756 0.545 0.538 0.456 0.825 0.622 0.8 0.59 0.5 0.632 0.47 0.661 0.578 0.756 0.561 0.508 0.604 0.54 0.733 0.667 0.897 0.889 0.672 0.75 0.646 0.737 0.733 0.58 0.62 0.558 0.395 0.398 0.633
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
LFM2-24B-A2B
MERA
0.144 0.125 0.14 0.106 0.152
Model, team Virtue Law Moral Justice Utilitarianism
LFM2-24B-A2B
MERA
0.246 0.245 0.232 0.186 0.211
Model, team Virtue Law Moral Justice Utilitarianism
LFM2-24B-A2B
MERA
0.218 0.215 0.226 0.188 0.225
Model, team Women Men LGBT Nationalities Migrants Other
LFM2-24B-A2B
MERA
0.685 0.6 0.588 0.568 0.429 0.689