Name: ERNIE 5.0
Author: Baidu

ERNIE 5.0

Baidu · 2026-01-22

ERNIE 5.0 is Baidu's closed, natively multimodal ultra-sparse Mixture-of-Experts foundation model, formally released on Qianfan on January 22, 2026 after a November 2025 preview; Baidu describes it as having 2.4 trillion parameters with less than 3% activated per token. It was trained from scratch across text, image, video, and audio under a unified autoregressive objective for multimodal understanding and generation, with modality-agnostic expert routing. The report describes training context extension to a rounded 128K tokens, while Qianfan's exact hosted profile reports a 248,832-token total context and text-only output for the ernie-5.0 endpoint; Qianfan also lists that endpoint as a deep-thinking model without an on/off control.

Benchmark scores

Benchmark	Score
AA-LCR	6.7
AA-Omniscience Index	-44.1
AIME 2025	85.0
Artificial Analysis Intelligence Index	21.9
Artificial Analysis Omniscience Accuracy	21.9
Artificial Analysis Omniscience Hallucination Rate	84.6
Buyout Game Bradley-Terry Rating	1043
Capability	140.1
CritPt	1.4
Extended NYT Connections 940-Puzzle Set	21.2
Extended NYT Connections Newest 100	20.2
GPQA Diamond	77.7
Humanity's Last Exam (Text-Only)	12.7
IFBench (Artificial Analysis)	41.4
LiveCodeBench (Artificial Analysis)	81.2
LLM Debate Benchmark	1283
LLM Persuasion Average Strength	0.8
LLM Persuasion Target Susceptibility	1.0
LLM Position Bias Decisive Pair Coverage	72.5
LLM Position Bias First-Option Lift	16.1
LLM Position Bias First-Shown Pick Rate	66.1
LLM Position Bias Order Flip Rate	45.0
LLM Position Bias Rating Bonus	0.3
LLM Sycophancy Conditional Rate	16.7
LLM Sycophancy Conditional Total Inconsistency	19.8
LLM Sycophancy Contrarian Rate	1.5
LLM Sycophancy Decisive Coverage	48.2
LLM Sycophancy Insufficient Answer Rate	37.2
LLM Sycophancy Rate	8.0
LLM Sycophancy Stripped Rate	8.0