Model

EXAONE 4.5

OPEN

LG AI Research · EXAONE family · released Apr 9, 2026

LG's first open-weight VLM, a fraction of the size of its 'K-EXAONE' sibling while matching its text performance; successor to EXAONE 4.0 32B already in corpus (dated 2025-07-15). Released for research/academic/educational use, not full commercial API.

ReasoningCodingVisionFunction callingTool useAgentic

1828.6

Elo · rank #78

Parameters

33B

Active params

33B (dense)

Context

256K tokens

Architecture

Vision-Language Model with proprietary vision encoder + hybrid-attention LLM backbone and multi-token prediction

License

EXAONE AI Model License (research/academic/educational use)

Languages

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision89.0%#21

best: Molmo 72B · 96.3%

AIMEMath92.6%#35

best: GPT-5.2 · 100.0%

CharXivVision71.7%#23

best: Muse Spark 1.1 (xhigh) · 88.4%

GPQA DiamondReasoning80.5%#77

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning11.6%#81

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning62.6%#30

best: MiniMax M3 · 83.0%

IFEvalReasoning89.6%#21

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding81.4%#43

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision75.2%#10

best: Seed 2.1 Pro · 92.6%

MathVistaVision85.0%#9

best: Seed 2.1 Pro · 90.7%

MMLU-ProKnowledge83.3%#37

best: Claude Fable 5 · 91.5%

MMMLUKnowledge85.4%#18

best: Gemini 3.1 Pro · 92.6%

MMMU-ProVision68.6%#31

best: Claude Opus 4.7 · 85.5%

MMMUVision78.7%#26

best: Claude Fable 5 · 89.3%

τ²-Bench AirlineAgents56.5%#6

best: Trinity-Large-Thinking · 88.0%

τ²-Bench RetailAgents77.9%#8

best: Claude Sonnet 4.6 · 91.7%

τ²-Bench TelecomAgents73.0%#28

best: Claude Opus 4.6 · 99.3%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

EXAONE family

Elo progression across releases

EXAONE 4.5Apr 20261828.6

Resources

HFLGAI-EXAONE/EXAONE-4.5-33B DOCShttps://github.com/LG-AI-EXAONE/EXAONE-4.5

API price weights · each benchmark row carries its own source badge (see methodology)