GigaChat3.5-432B-A28B-Reasoning
GigaChat
Создан
03.09.2026 19:53
0.423
Общий результат
13
Место в рейтинге
В топе по задачам:
5
SAGE
6
RussianRegions
7
RUBIN
9
IFHardBench
10
Riddles
Задачи с минимальными баллами:
19
Characters
18
SOB-Hard
18
GorillaHard
16
Humor
16
Enantiosemy
Модель, команда
Результат
{{ formatValue(getLeaderboardPlace(submit)) }}
{{ submit.name }}
{{ transformSize(submit.model_size) }}
{{ submit.team_name }}
{{ formatValue(getResultValue(submit)) }}
| {{ column.title }} | Время |
|---|---|
| {{ formatValue(getColumnValue(submit, column)) }} | {{ formatValue(getTotalTime(submit)) }} |
Все задачи
| Задача | Результат | Метрика |
|---|---|---|
| SAGE | 0.684 / 0.691 / 0.688 |
PUNCT_F1
SPELL_F1
ERRANT_F1
|
| Humor | 0.282 / 0.301 |
Exact match
LLM judge score
|
| LIMUR | 0.626 / 0.629 |
Exact match
LLM judge score
|
| RUBIN | 0.697 / 0.704 |
Exact match
LLM judge score
|
| Riddles | 0.492 / 0.511 |
Exact match
LLM judge score
|
| SOB-Hard | 0.116 / 0.218 / 0.744 / 0.184 / 0.223 / 0.751 |
balance_score
task_pass_rate
format_pass_rate
sample_pass_rate
content_pass_rate
constraint_pass_rate
|
| Characters | 0.523 / 0.578 |
Exact match
LLM judge score
|
| Enantiosemy | 0.348 / 0.514 |
Exact match
LLM judge score
|
| GorillaHard | 0.467 / 0.8 / 0.329 / 0.55 / 0.567 / 0.857 / 0.324 / 0.402 / 1 / 0.052 / 0.982 / 0.865 / 0.074 / 1 |
balance_score
clarify_recall
args_match_rate
tool_match_rate
dialog_pass_rate
format_pass_rate
sample_pass_rate
abstention_recall
cost_optimal_rate
false_clarify_rate
constraint_pass_rate
tool_in_catalog_rate
false_abstention_rate
injection_resistance_rate
|
| IFHardBench | 0.846 / 0.467 / 0.821 |
balance_score
sample_pass_rate
constraint_pass_rate
|
| NewReasoning | 0.6 / 0.639 |
Exact match
LLM judge score
|
| RussianRegions | 0.184 / 0.658 / 0 / 0.368 |
Exact match
LLM judge score
Group Exact match
Group Judge Score
|
Информация о сабмите
Команда:
GigaChat
Название ML-модели:
GigaChat3.5-432B-A28B-Reasoning
Ссылка на ML-модель:
https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-ReasoningРазмер модели
432.0B
Тип модели:
Открытая
GigaChat 3.5 Reasoning is the first GigaChat model with full reasoning trained with online RL. Compared with GigaChat 3.5 Ultra Instruct, the largest gains are in mathematics, code, instruction following, and structured output. It uses a custom hybrid architecture that combines Multi-head Latent Attention (MLA) with GatedDeltaNet linear-attention layers.
Лицензия:
MIT
Версия MERA
v2.0.0
Версия Torch
2.13.0
Версия кодовой базы
9efaf1d
Версия CUDA
None
Precision весов модели
auto
Сид
1234
Батч
1
Версия transformers
5.16.1
Количество GPU и их тип
-
Архитектура
gigachat-openai
mera@a-ai.ru
GitHub
hugging face