ERNIE X1
Last reviewed
Jun 8, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,551 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,551 words
Add missing citations, update stale details, or suggest a clearer explanation.
ERNIE X1 is a deep-reasoning large language model developed by Baidu, the Chinese search and artificial-intelligence company, as part of its ERNIE (Wenxin) family. Announced on March 16, 2025, ERNIE X1 was Baidu's first model in the deliberate-reasoning, or "deep thinking," category, designed to produce explicit chains of thought before answering and to call external tools while solving problems.[1][2] Baidu launched it alongside the multimodal foundation model ERNIE 4.5 and, in the same announcement, made its consumer chatbot ERNIE Bot free ahead of schedule.[1] The company positioned X1 as a direct competitor to DeepSeek-R1 and the OpenAI o-series, claiming performance on par with DeepSeek-R1 at roughly half the price.[1][3] A successor, ERNIE X1.1, followed on September 9, 2025.[4]
ERNIE X1 belongs to the class of reasoning models, systems that allocate additional inference-time computation to internal reasoning steps rather than answering immediately. Where Baidu's general-purpose ERNIE models prioritize fast responses across a broad range of tasks, the X1 line is tuned for problems that benefit from multi-step deliberation, such as mathematics, logic, and code. Baidu describes X1 as a multimodal deep-thinking model with built-in tool use, able to plan, reflect, and revise its own reasoning.[2]
The model is part of Baidu's broader ERNIE (Enhanced Representation through kNowledge IntEgration), known in Chinese as Wenxin, series. Within that series, the numeric line (for example ERNIE 4.5 and ERNIE 5.0) provides the general multimodal foundation, while the "X" line denotes the reasoning-specialized variants built on top of that foundation.
| Specification | Detail |
|---|---|
| Developer | Baidu |
| Family | ERNIE (Wenxin) |
| Model type | Reasoning-focused large language model ("deep thinking") |
| Announced | March 16, 2025[1] |
| Launched alongside | ERNIE 4.5 |
| Modalities | Multimodal (text, with image understanding and generation via tools)[2] |
| Tool use | Yes: advanced search, document Q&A, image understanding, AI image generation, code interpreter, webpage reading[2] |
| Positioning | Comparable to DeepSeek-R1 at about half the price (Baidu claim)[1][3] |
| API price (X1) | As low as RMB 0.002 per 1,000 input tokens; RMB 0.008 per 1,000 output tokens (Baidu)[2] |
| Availability | ERNIE Bot (free); Qianfan / Baidu AI Cloud API[2] |
| Successor | ERNIE X1.1 (September 9, 2025)[4] |
Baidu has developed the ERNIE family of language models since 2019, originally as pretraining methods that integrated knowledge-graph information into transformer-based representations. The series grew into a line of large generative models, and in March 2023 Baidu launched ERNIE Bot (Wenxiaoyan), one of the first Chinese answers to OpenAI's ChatGPT. Despite being an early entrant, ERNIE Bot was widely reported to have struggled to achieve broad adoption against newer rivals.[3]
The competitive landscape shifted sharply in early 2025 when the Chinese startup DeepSeek released DeepSeek-R1, a low-cost reasoning model that drew global attention and pressured incumbents on both capability and price. ERNIE X1 was Baidu's response: a reasoning model intended to match R1's quality while undercutting it on cost. ERNIE X1 and ERNIE 4.5 were built to share Baidu's foundation-model technology, with X1 adding the reasoning, planning, and tool-use layers on top.[2]
Baidu unveiled ERNIE 4.5 and ERNIE X1 on March 16, 2025.[1][2] In the same announcement, the company made ERNIE Bot freely available to individual users ahead of its previously scheduled date, a move widely read as a reaction to the free availability of competing chatbots in China.[1][3]
Pricing was central to the launch. Baidu said ERNIE X1 delivered performance "on par with DeepSeek-R1 at only half the price," with API rates as low as RMB 0.002 per 1,000 input tokens and RMB 0.008 per 1,000 output tokens.[2] For comparison, Baidu priced the ERNIE 4.5 foundation model at RMB 0.004 per 1,000 input tokens and RMB 0.016 per 1,000 output tokens, and claimed it outperformed GPT-4.5 on several benchmarks at roughly 1% of GPT-4.5's price.[2] These claims were the company's own and were not independently verified at launch; the DeepSeek-R1 comparison in particular reflects Baidu's marketing rather than a neutral evaluation.
The launch intensified an ongoing price war among Chinese model providers, including DeepSeek, Qwen (from Alibaba), and Kimi K2 (from Moonshot AI), in which providers repeatedly cut token prices and, in several cases, open-sourced models to compete for developers. In April 2025 Baidu teased lower-cost "Turbo" variants of both models, ERNIE X1 Turbo and ERNIE 4.5 Turbo, promising faster and cheaper performance.[5]
Baidu describes ERNIE X1 as a deep-thinking reasoning model with "enhanced capabilities in understanding, planning, reflection, and evolution."[2] Rather than producing an immediate answer, the model generates an internal chain of reasoning, which Baidu credits for stronger results on tasks involving logic, mathematics, and calculation.[2] Baidu highlighted particular strengths in Chinese-language tasks, including Chinese question answering, literary creation, manuscript writing, and dialogue.[2]
A defining feature of X1 is native tool use. The model can autonomously invoke a range of tools, including advanced search, document question answering, image understanding, AI image generation, a code interpreter, webpage reading, and various specialized search functions.[2] This agentic, tool-calling behavior distinguishes X1 from a pure text generator and aligns it with the broader industry move toward reasoning models that combine internal deliberation with external actions, exemplified by OpenAI o1, OpenAI o3, and DeepSeek-R1.
At launch, Baidu's capability claims for ERNIE X1 were stated comparatively (parity with DeepSeek-R1) rather than through a detailed published benchmark suite, and specific scores were not provided in the initial release.[3] Independent benchmark figures attributable to a neutral evaluator were not available at the time of writing, so the performance characterization here reflects Baidu's own statements.
Baidu released an upgraded reasoning model, ERNIE X1.1, on September 9, 2025, at its WAVE SUMMIT 2025 developer conference in Beijing.[4] The company reported gains over the original ERNIE X1 of 34.8% in factuality, 12.5% in instruction following, and 9.6% in agentic capabilities.[4] Baidu said X1.1 was built on the ERNIE 4.5 multimodal foundation using an iterative hybrid reinforcement-learning framework that combined mixed reinforcement learning with iterative self-distillation.[4]
Baidu also claimed that ERNIE X1.1 surpassed DeepSeek-R1-0528 in overall performance and was on par with top-tier models such as GPT-5 and Gemini 2.5 Pro, though it did not publish the underlying benchmark detail; these remain company claims.[4] ERNIE X1.1 was made available through the ERNIE Bot website and the Wenxiaoyan app, and was deployed on Baidu's Qianfan platform (the Model-as-a-Service offering of Baidu AI Cloud) for enterprise clients and developers.[4]
Alongside ERNIE X1.1, Baidu open-sourced ERNIE-4.5-21B-A3B-Thinking, a mixture-of-experts reasoning model with 21 billion total parameters and about 3 billion active parameters and a 128K-token context window, released on Hugging Face and Baidu AI Studio.[4] This followed Baidu's earlier decision, on June 30, 2025, to open-source the broader ERNIE 4.5 model family, a set of variants ranging from a 0.3 billion-parameter dense model up to a 424 billion-parameter MoE model, released under the Apache License 2.0 across Hugging Face, GitHub, and the PaddlePaddle ecosystem.[6][7] These open releases concerned the ERNIE 4.5 line; the X1 reasoning models themselves were offered primarily through ERNIE Bot and the Qianfan API rather than as fully open weights, except for the smaller "Thinking" variant noted above.
Baidu continued to advance the wider ERNIE series after X1. At Baidu World 2025 on November 13, 2025, the company unveiled ERNIE 5.0, a natively omni-modal foundation model that jointly handles text, images, audio, and video, made available in preview through ERNIE Bot and to enterprises through Qianfan.[8] Some reports described ERNIE 5.0 as a roughly 2.4 trillion-parameter system, though Baidu's official announcement did not state a parameter count.[8]
ERNIE X1 marked Baidu's entry into the reasoning-model race that DeepSeek-R1 had thrust to the center of attention in early 2025. By pairing a reasoning model with aggressive pricing and free consumer access to ERNIE Bot, Baidu helped accelerate a price war among China's leading model developers and reinforced a competitive dynamic in which Chinese firms increasingly released or open-sourced capable models at low cost. The X1 line, continued through X1.1 and complemented by the open-sourced ERNIE 4.5 family and the later ERNIE 5.0 foundation model, illustrates how rapidly Chinese providers iterated on reasoning capabilities during 2025, narrowing the gap with leading models from OpenAI and Google while competing chiefly on cost and accessibility.