Model
Model explorer

LG EXAONE Deep 32B

OPEN
LG AI Research · EXAONE Deep family · released Mar 18, 2025

Reasoning-specialized <thought>-tag CoT model rivaling much larger models on AIME/MATH.

ReasoningCodingVisionFunction callingTool useAgentic
1315.7
Elo · rank #167
Parameters
32B
Active params
32B (dense)
Context
32K tokens
Architecture
Dense Transformer, 64 layers, GQA (40 Q-heads / 8 KV-heads), <thought>-tag CoT reasoning
License
EXAONE AI Model License Agreement 1.1 - NC
Languages
2+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath65.8%#110
best: GPT-5.2 · 100.0%
GPQA DiamondReasoning66.1%#145
best: GPT-6 Astra · 96.0%
LiveCodeBenchCoding59.5%#99
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math95.7%#32
best: GPT-5 · 99.4%
MMLU-ProKnowledge74.0%#80
best: Claude Fable 5 · 91.5%
MMLUKnowledge83.0%#74
best: OpenAI o3 · 92.9%
Run it locally
VRAM @ Q4
20 GB
VRAM @ FP16
62 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
GGUF Q4 · AWQ
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 Deep family
Elo progression across releases
LG EXAONE Deep 32BMar 20251315.7
Resources
HFLGAI-EXAONE/EXAONE-Deep-32BGHLG-AI-EXAONE/EXAONE-Deep
API price weights · each benchmark row carries its own source badge (see methodology)