跳轉到主要內容

EXAONE 4.5 33B (Reasoning)

LG AI ResearchProprietary

描述

EXAONE 4.5 is the first open-weight vision-language model from LG AI Research. It is a dense 33B multimodal model (31.7B language + 1.29B vision encoder) built by adding a dedicated visual encoder to the EXAONE 4.0 framework. It supports a 262,144-token context window, hybrid reasoning (thinking on by default), and strong performance on document understanding and Korean contextual reasoning while retaining EXAONE language capabilities. License: EXAONE AI Model License Agreement 1.2 - NC. Knowledge cutoff: December 2024.

發布日期
2026-04-09
參數規模
33.0B
上下文長度
—
支援模態
—

能力雷達圖

13
general
24
coding
79
reasoning
59
science
70
agents
80
multimodal

排行榜排名

領域#排名分數來源
程式碼能力榜375
38.0
AA
通用能力榜283
45.0
AA
多模態榜71
53.0
LS
科學能力272
50.0
AA

基準測試分數 (LLM Stats)

(LLM Stats (zeroeval))

Chat

IFEvalGoogle Research (2023)89.6%自報
Tau2 Airline56.5%自報

Communication

Tau2 Retail77.9%自報
Tau2 Telecom73.0%自報

Language

MMLU-Pro83.3%自報

Math

AIME 202592.9%自報
AIME 202692.6%自報
MathVista-Mini85.0%自報

Multimodal

MMMU78.7%自報

Reasoning

LiveCodeBench v681.4%自報
GPQANYU + Cohere + Anthropic (2023)80.5%自報
AA-LCR50.6%自報

Vision

AI2D89.0%自報
MMStar74.9%自報
MMMU-Pro68.6%自報

AA 評測指數

(Artificial Analysis)
Gpqa(NYU + Cohere + Anthropic (2023))
79.4
Tau2(Sierra + U Toronto + Vector Institute (2025))
78.1
Ifbench(Google Research (2023))
58.0
Lcr(Artificial Analysis)
54.7
Coding Index(Artificial Analysis)
23.6
Terminalbench V2 1
21.3
Terminalbench Hard(Stanford × Laude Institute (2026))
20.5
Intelligence Index(Artificial Analysis)
13.2
Hle(Center for AI Safety + Scale AI (2025))
12.9

LLM Stats 分類評分

(LLM Stats (zeroeval))
Instruction Following
90
Math
90
Structured Output
90
Language
80
Legal
80
Multimodal
80
Physics
80
Reasoning
80
Finance
80
General
80
Healthcare
80
Biology
80
Chemistry
80
Vision
80
Chat
70
Communication
70
Tool Calling
70
Long Context
50

定價

輸入價格免費
輸出價格免費
混合價格(3:1)免費

速度

Tokens/秒0.0
首Token延遲0.00s
首回答延遲0.00s

供應商價格排行

暫無提供商資料

外部連結