Model

ERNIE 4.5 300B-A47B

OPEN

Baidu · ERNIE 4.5 family · released Jun 30, 2025

Flagship open-weights text MoE anchoring the 0.3B-424B ERNIE 4.5 family; scores below the median of comparable open models on Artificial Analysis's Intelligence Index.

ReasoningCodingVisionFunction callingTool useAgentic

1445.5

Elo · rank #132

Parameters

300B

Active params

47B (MoE)

Context

128K tokens

Architecture

Mixture-of-Experts, 64 text experts / 8 active per token, 54 layers, GQA 64Q/8KV

License

Apache-2.0

Languages

—

API price (in/out)

$0.28 / $1.1

Modalities

text

Benchmark results

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

AIMEMath35.1%#131

best: GPT-5.2 · 100.0%

BIG-Bench HardReasoning94.3%#1

best: this model · 94.3%

C-EvalKnowledge90.6%#12

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge90.2%#2

best: Doubao-1.5-Pro · 90.9%

DROPReasoning91.1#5

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning81.0%#74

best: GPT-5.6 · 94.6%

GSM8KMath96.6%#3

best: Llama 3.1 405B · 96.8%

Humanity's Last ExamReasoning4.0%#127

best: Claude Sonnet 5 · 57.4%

HumanEval+Coding92.1%#5

best: Mistral Small 3.2 (24B) · 92.9%

IFBenchReasoning39.0%#47

best: MiniMax M3 · 83.0%

IFEvalReasoning88.0%#34

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding38.8%#133

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

MATH-500Math96.4%#24

best: GPT-5 · 99.4%

MBPP+Coding81.0%#6

best: Llama 3.1 405B · 88.6%

MMLU-ProKnowledge78.4%#62

best: Claude Fable 5 · 91.5%

MMLUKnowledge86.5%#41

best: OpenAI o3 · 92.9%

SimpleQAKnowledge45.9%#5

best: GPT-4.5 · 62.5%

Run it locally

VRAM @ Q4

151 GB

VRAM @ FP16

603 GB

Fits on (Q4)

M3 Ultra 512GBB200 192GB

300B total / 47B active MoE far exceeds a single RTX 4090's 24GB even at 4-bit; needs multi-GPU/datacenter hardware.

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-4.5-300B-A47B-PT GHPaddlePaddle/ERNIE APIAvailable via Baidu API

API price $0.28/$1.1 · each benchmark row carries its own source badge (see methodology)