Model

ERNIE 4.5 VL 424B-A47B

OPEN

Baidu · ERNIE 4.5 family · released Jun 30, 2025

Largest open ERNIE 4.5 release; supports thinking/non-thinking modes, with Baidu claiming near-o1 results on MathVista, MMMU and VisualPuzzle.

ReasoningCodingVisionFunction callingTool useAgentic

1354.6

Elo · unrated

Parameters

424B

Active params

47B (MoE)

Context

128K tokens

Architecture

Multimodal heterogeneous MoE, 128 experts (64 text + 64 vision), 8 active per modality, 54 layers

License

Apache-2.0

Languages

—

API price (in/out)

$0.42 / $1.25

Modalities

text · vision · video

Benchmark results

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

AI2DVision95.3%#6

best: Molmo 72B · 96.3%

ChartQAVision86.8%#17

best: MiniMax-VL-01 · 91.7%

DocVQAVision93.1%#23

best: Qwen2-VL-72B · 96.5%

MathVistaVision78.9%#15

best: Seed 2.1 Pro · 90.7%

MMBench (Chinese)Vision90.9%#1

best: this model · 90.9%

MMBench (English)Vision92.0%#6

best: Qwen3.5-397B-A17B · 93.7%

MMMU-ProVision58.8%#39

best: Claude Opus 4.7 · 85.5%

MMMUVision70.0%#55

best: Claude Fable 5 · 89.3%

OCRBenchVision872#6

best: InternVL3-78B · 906

Run it locally

VRAM @ Q4

245 GB

VRAM @ FP16

979 GB

Fits on (Q4)

M3 Ultra 512GB

424B total / 47B active multimodal MoE needs roughly 1TB VRAM at FP16 and stays far beyond a single RTX 4090 even at 4-bit.

Quantizations

GGUF Q4 · FP8 · INT4 (W4A16) · 2-bit

ERNIE 4.5 family

Elo progression across releases

ERNIE 4.5 300B-A47BJun 20251445.5 ERNIE 4.5 VL 424B-A47BJun 20251354.6

Resources

HFbaidu/ERNIE-4.5-VL-424B-A47B-PT GHPaddlePaddle/ERNIE APIAvailable via Baidu API

API price $0.42/$1.25 · each benchmark row carries its own source badge (see methodology)