ERNIE X1

What ERNIE X1 is

ERNIE X1 is a reasoning model from Baidu, released on 16 March 2025 together with the ERNIE 4.5 model^[1]. Baidu describes it as a deep-thinking reasoning model with enhanced capabilities in understanding, planning, reflection and evolution^[1]. In practice, a "deep-thinking" or reasoning model is one trained to spend more computation working through a problem before it gives an answer, rather than producing a reply in a single pass. Baidu also calls X1 its first multimodal reasoning model capable of tool use^[1]: multimodal means it can handle more than plain text, such as images, and tool use means it can call external functions instead of relying only on what it absorbed during training.

Release

The model was announced on 16 March 2025 in the same release as ERNIE 4.5^[1]. The headline of Baidu's launch announcement said the company had made ERNIE Bot free ahead of schedule, alongside ERNIE 4.5 and X1^[1]. At launch, X1 was free to individual users on the ERNIE Bot website, while access through an API on Baidu's Qianfan platform was announced as coming soon^[1][3].

How Baidu says it works

Baidu named three key technologies behind X1. The first is a Progressive Reinforcement Learning Method^[1]. Reinforcement learning is a training approach in which a model is rewarded for outputs that satisfy some criterion, and gradually adjusts its behaviour to earn more of that reward; the word "progressive" suggests the criteria are tightened over the course of training. The second is an end-to-end training approach that integrates chains of thought and action^[1]. A chain of thought is a step-by-step line of reasoning written out before the final answer, while "action" refers to the model doing something in the world, such as calling a tool; training the two together is meant to keep reasoning and tool use in one loop rather than as separate stages. The third is a unified multi-faceted reward system^[1], meaning the reward that drives training is combined from several different criteria rather than a single score.

Strengths and supported tools

Baidu cites Chinese knowledge question answering, literary creation, manuscript writing, dialogue, logical reasoning and complex calculations as the model's strengths^[1]. It also lists the tools X1 could use: advanced search, document question answering, image understanding, AI image generation, a code interpreter, webpage reading, TreeMind mapping, Baidu academic search, business information search and franchise information search^[1]. These cover both retrieval — finding material on the web or in documents — and generation, such as producing images or running code.

Baidu did not publish specific benchmark scores for X1 at launch^[2], so its capability claims rested on the company's own descriptions rather than on numbers that outside parties could check against other models.

Pricing

Baidu claimed performance on par with DeepSeek-R1 at half the price, with prices as low as RMB 0.002 per thousand input tokens and RMB 0.008 per thousand output tokens^[1]. Token-based pricing of this kind charges by the amount of text processed, so input tokens are the prompt sent to the model and output tokens are what it generates.

Later versions

On 25 April 2025 Baidu released ERNIE X1 Turbo at half of X1's price^[4]. In that release the company described ERNIE X1, which had debuted on 16 March 2025, as a predecessor of the Turbo models launched that day^[4]. On 9 September 2025 Baidu released ERNIE X1.1, which it compared against X1^[5].

Where it sits

X1 sits in the wave of reasoning models that followed the demonstration that extra computation at answer time could improve results on hard problems. Its distinguishing choices were breadth of built-in tools, an emphasis on Chinese-language tasks, and aggressive pricing^[1]. Because Baidu released no benchmark figures at launch^[2], and because the model was revised twice within the same year — into X1 Turbo in April and X1.1 in September^[4][5] — X1 is best understood as the first entry in a line of Baidu reasoning models rather than a fixed reference point.