ERNIE X1.1

What ERNIE X1.1 is

ERNIE X1.1 is a reasoning model from Baidu, unveiled on 9 September 2025 at WAVE SUMMIT 2025, the company's deep-learning developer conference^[1][2]. It is an update to ERNIE X1, and Baidu frames it around three things: giving factual answers, following instructions, and carrying out tasks that need several steps of action — what the company calls agentic capabilities^[2]. Ordinary users reach it through the ERNIE Bot website at ernie.baidu.com and through the Wenxiaoyan app, while enterprises and developers get it through deployment on Baidu's Qianfan platform^[2].

How it was trained

The model is built on ERNIE 4.5^[2]. Baidu describes the training method as an iterative hybrid reinforcement-learning framework that combines mixed reinforcement learning with iterative self-distillation^[2]. Reinforcement learning, in general, means letting a model produce answers, scoring those answers with a reward signal, and pushing the model toward the outputs that score well. Calling the setup "mixed" indicates that more than one kind of reinforcement-learning objective is used together rather than a single one. Self-distillation means feeding a model's own better outputs back in as training targets for the next round, and "iterative" means that loop runs repeatedly, so each round starts from the improved version left by the previous one. The point of combining the two is to keep improving the model without needing a fresh, separately labelled dataset at every stage.

Reported gains and comparisons

Measured against ERNIE X1, Baidu reported factuality up 34.8%, instruction following up 12.5% and agentic capabilities up 9.6%^[2]. The company also said the model's overall performance surpassed DeepSeek-R1-0528 on multiple benchmarks and was on par with GPT-5 and Gemini 2.5 Pro^[2]. These are the vendor's own comparisons: benchmark results depend on which tests are chosen and how they are run, so they describe Baidu's claims rather than an independent ranking.

What shipped alongside it

ERNIE X1.1 was announced together with PaddlePaddle v3.2, the next version of Baidu's open-source deep-learning framework, and with an open-source model called ERNIE-4.5-21B-A3B-Thinking^[2]. That open model supports a 128K context window — the amount of text it can take in and keep in view at one time — and is available on Hugging Face and Baidu AI Studio^[1].

The same announcements carried some numbers about Baidu's training and developer base. Baidu reported a pre-training MFU of 47% on its ERNIE-4.5-300B-A47B model^[1]; MFU, or model FLOPs utilization, is the share of a chip's peak arithmetic throughput that training actually puts to use, so it is a measure of training efficiency rather than of model quality. The company said the PaddlePaddle-ERNIE ecosystem had served 23.33 million developers and 760,000 enterprises^[1], and that 45% of new code written inside Baidu is now generated by AI^[1].

Where it sits

ERNIE X1.1 is best read as a step in Baidu's reasoning line rather than a standalone launch. It sits on top of the ERNIE 4.5 base, it arrives in the same breath as a framework release and an open-weight sibling, and it is distributed through both consumer apps and a cloud platform^[1][2]. That arrangement — a base model family, a training framework, an open release for outside developers, and a hosted service for companies — is how Baidu presents the model as part of a stack rather than a single product. The capability Baidu emphasizes most, agentic behaviour, is also the one where evaluation is least settled: a model that plans and takes actions in several steps is harder to score than one that answers a single question, which is why the company's percentage gains are a starting point for judging the release rather than a conclusion.