Qwen3.8 Max (0902)
AlibabaQwenОткрытые весаQwen3.8-Max License · Коммерческое использование
Описание
Qwen3.8 Max, released as the Qwen3.8-2.4T-A95B open-weight checkpoint, is Qwen's flagship mixture-of-experts model for coding, research, professional workflows, and long-horizon agents. It has 2.4 trillion total parameters with 95 billion active parameters, a native 262,144-token context window extensible to about 1 million tokens, and always-on thinking. QwenCloud serves the same model in the Qwen3.8 Max API family with text and image input.
Дата выхода
2026-09-02
Параметры
2.4T
Длина контекста
1.0M
Модальности
image, pdf, text, video
Радар способностей
45
general
72
coding
93
reasoning
69
science
60
agents
90
multimodal
Рейтинги
| Домен | #Место | Оценка | Источник |
|---|---|---|---|
| Агентные возможности | 18 | 61.0 | LS |
| Рейтинг кодинга | 37 | 92.0 | AA |
| Общий рейтинг | 28 | 78.0 | AA |
| Математическое мышление | 27 | 91.0 | LB |
| Мультимодальный рейтинг | 1 | 84.0 | LS |
| Рассуждения | 21 | 88.0 | LB |
| Наука | 56 | 80.0 | AA |
Оценки бенчмарков (LLM Stats)
(LLM Stats (zeroeval))Agents
QwenSVG
1713.00 / 2000Сам.
AndroidWorld
85.3%Сам.
MobileWorld
77.8%Сам.
AndroidBench
75.1%Сам.
CoWorkBench
74.8%Сам.
Toolathlon
72.5%Сам.
Workspace Bench
67.7%Сам.
Job Bench
53.4%Сам.
OneMillion Bench
52.5%Сам.
Agents' Last Exam
52.4%Сам.
MLS-Bench Lite
41.0%Сам.
AutomationBench
27.3%Сам.
Code
QwenReactBench
1724.00 / 2000Сам.
PaperBench
93.0%Сам.
QwenSWEBench
80.7%Сам.
FrontierSWE
73.5%Сам.
SkillsBench
70.2%Сам.
Vision2Web
69.0%Сам.
QwenQoderBench
58.4%Сам.
DeepSWE 1.1
56.6%Сам.
NL2Repo
55.9%Сам.
Healthcare
HealthBench
60.2%Сам.
Instruction Following
IFBench
82.8%Сам.
Knowledge
PLawBench
73.2%Сам.
PRBench-Finance
58.3%Сам.
PRBench-Legal
57.6%Сам.
Long Context
MRCR v2 (8-needle)
92.9%Сам.
LongBench v2
66.3%Сам.
Multimodal
OSWorld-Verified
86.1%Сам.
Reasoning
GPQANYU + Cohere + Anthropic (2023)
92.6%Сам.
Terminal-Bench 2.1
86.6%Сам.
SWE-Bench ProPrinceton NLP (2024)
67.7%Сам.
Humanity's Last Exam (with tools, text-only)
56.2%Сам.
Humanity's Last Exam
43.6%Сам.
Search
WideSearch
81.9%Сам.
Vision
VideoMME w sub.
90.4%Сам.
RealWorldQA
88.0%Сам.
ScreenSpot Pro
84.5%Сам.
MMMU-Pro
82.3%Сам.
LVBench
81.8%Сам.
ERQA
77.8%Сам.
PerceptionBench
63.5%Сам.
Индексы оценки AA
(Artificial Analysis)Gpqa(NYU + Cohere + Anthropic (2023))92.8
Terminalbench V2 188.8
Lcr(Artificial Analysis)80.3
Coding Index(Artificial Analysis)76.2
Scicode(UIUC + Argonne National Lab (2024))52.1
Tau Banking47.8
Intelligence Index(Artificial Analysis)45.4
Hle(Center for AI Safety + Scale AI (2025))43.1
Оценки категорий LLM Stats
(LLM Stats (zeroeval))Physics90
Biology90
Chemistry90
Video90
Instruction Following80
Multimodal80
Search80
Spatial Reasoning80
Grounding80
Vision80
Long Context70
Reasoning70
Structured Output70
General70
Agents70
Code70
Productivity60
Healthcare60
Tool Calling60
Math50
Цены
Цена ввода$2 / 1M токенов
Цена вывода$6 / 1M токенов
Смешанная цена (3:1)$3 / 1M токенов
Цена чтения кэша$0.25 / 1M токенов
Цена записи кэша$2.5 / 1M токенов
Скорость
Токенов/сек37.8
Задержка первого токена1.95s
Время до первого ответа54.86s
Рейтинг цен провайдеров
Рейтинг цен провайдеров
6 провайдеров
Самый дешевый: DeepInfraСамый дорогой: EmpirioLabs AI
ПровайдерВводВывод
1DeepInfraСамый дешевый
$0
$0
2Fireworks
$0
$0.00001
3Novita
$0
$0.00001
4Together
$0
$0.00001
5AlibabaОсновной
$2
$6
6EmpirioLabs AI
$2
$6
Сравнение цен разных API-провайдеров для этой модели.