Model

LG EXAONE 4.0 32B (Non-reasoning)

OPEN

LG AI Research · EXAONE 4.0 family · released Jul 15, 2025

Default (non-reasoning) mode of Korea's first open-weight hybrid model; faster, lower scores on math/coding than reasoning mode.

ReasoningCodingVisionFunction callingTool useAgentic

1211.0

Elo · rank #190

Parameters

32B

Active params

32B (dense)

Context

131K tokens

Architecture

Dense hybrid-attention transformer, 64 layers, local:global attention 3:1, GQA (40Q/8KV), QK-Reorder-Norm; single checkpoint with toggleable non-reasoning/reasoning modes

License

EXAONE AI Model License Agreement 1.2 (NC)

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

AIMEMath35.9%#131

best: GPT-5.2 · 100.0%

BFCL v3Agents65.2%#20

best: Hunyuan-A13B · 78.3%

GPQA DiamondReasoning63.7%#160

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning4.9%#128

best: Claude Opus 5 · 64.7%

IFBenchReasoning33.5%#68

best: MiniMax M3 · 83.0%

IFEvalReasoning84.8%#63

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding43.3%#131

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

MMLU-ProKnowledge77.6%#68

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge89.8%#23

best: Qwen3.7-Max · 95.0%

MMMLUKnowledge80.6%#24

best: Gemini 3.1 Pro · 92.6%

tau-benchAgents55.9%#16

best: Claude Opus 4.1 · 82.4%

Run it locally

VRAM @ Q4

20 GB

VRAM @ FP16

64 GB

Fits on (Q4)

RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

FP8 · GGUF Q4

Fine-tune it

Research only

QLoRA22.2 GB1× RTX 3090 24GB

LoRA68.6 GB1× A100 80GB

Full fine-tune514.0 GB4× H200 141GB

QLoRA SFT on ~10k samples ≈ $16.48 (1× RTX 3090 24GB)

Plan a fine-tune →

EXAONE 4.0 family

Elo progression across releases

LG EXAONE 4.0 32B (Non-reasoning)Jul 20251211.0 LG EXAONE 4.0 32B (Reasoning)Jul 20251551.7

Resources

HFLGAI-EXAONE/EXAONE-4.0-32B GHLG-AI-EXAONE/EXAONE-4.0

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