Baidu unveils ERNIE X1.1 at WAVE SUMMIT 2025

Background

ERNIE X1 had been Baidu's reasoning model since March 2025, and ERNIE X1 Turbo followed in April. ERNIE X1.1 was the next version in that line, presented as an upgrade of X1 in several capabilities rather than as a new family. Reasoning models are judged on several dimensions: whether their answers are factually accurate, how closely they follow instructions, and how well they handle agentic tasks, in which a model plans steps and uses tools to reach a goal.

What happened

Baidu unveiled ERNIE X1.1 on 9 September 2025 at WAVE SUMMIT 2025, the company's deep-learning developer conference.^[1][2] The company presented it as a reasoning model with upgrades in its key capabilities.^[1]

According to Baidu, factuality rose 34.8% compared with ERNIE X1, instruction following rose 12.5% and agentic capabilities rose 9.6%.^[2] Baidu said X1.1 surpassed DeepSeek-R1-0528 and was on par with GPT-5 and Gemini 2.5 Pro.^[2]

Key details

ERNIE X1.1 was trained on ERNIE 4.5 using an iterative hybrid reinforcement-learning framework.^[2] Reinforcement learning trains a model by rewarding answers that score well, and an iterative approach repeats this cycle, using the results of each round to guide the next one.

At the same conference Baidu also open-sourced ERNIE-4.5-21B-A3B-Thinking and launched PaddlePaddle v3.2.^[2] Baidu said the PaddlePaddle-ERNIE ecosystem had served 23.33 million developers and 760,000 enterprises.^[2]

Impact

X1.1 was announced as an upgrade within Baidu's reasoning-model line and placed against both DeepSeek-R1 and GPT-5 and Gemini 2.5 Pro in Baidu's own comparisons.^[2] Because the figures in the announcement are Baidu's own, readers should treat the percentages as the company's claims. The announcement also paired the model with an open-source release and a framework update, so the model was presented as part of a wider developer ecosystem rather than as a standalone product.

Factuality refers to how often a model's statements match checkable facts, while instruction following measures how faithfully it carries out the requirements in a prompt.