Baidu drops ERNIE X1.1 reasoning model to take on GPT-5

Baidu is making a major push into advanced reasoning, taking direct aim at the top frontier labs.

At its WAVE SUMMIT 2025 conference on September 9, Baidu introduced ERNIE X1.1, its latest reasoning model. The company claims the new model edges past DeepSeek R1-0528 in overall performance and matches top-tier heavyweights like GPT-5 and Gemini 2.5 Pro.

Alongside the flagship model, Baidu unveiled an open-source reasoning model, updated its core training framework, and revamped its AI coding tools.

Less hallucination and an open-source MoE

ERNIE X1.1 focuses heavily on reliability. Compared to its predecessor ERNIE X1, Baidu reports a 34.8% jump in factuality, a 12.5% improvement in instruction following, and a 9.6% boost in agentic capabilities.

Built on top of ERNIE 4.5, X1.1 incorporates an iterative hybrid reinforcement learning system combining mixed RL and self-distillation. It is already accessible via the ERNIE Bot site, the Wenxiaoyan app, and Baidu's Qianfan cloud platform.

At the same time, Baidu released ERNIE-4.5-21B-A3B-Thinking under an open-source license on Hugging Face and Baidu AI Studio. This follows Baidu's open-sourcing of 10 other ERNIE 4.5 models on June 30.

The new open model features a 128K context window. As a sparse Mixture-of-Experts (MoE) model with 21 billion total parameters and just 3 billion active parameters, it aims to deliver near frontier-level math and logic reasoning while running on far leaner hardware.

Multi-agent coding and 47% training efficiency

Baidu also updated its developer stack. The new Baidu Comate 3.5S coding assistant shifts from single-agent tasks to multi-agent collaboration, dynamically spinning up multiple sub-agents to tackle complex engineering tasks. It also integrates Rules and MCP.

According to Baidu, Comate has served over 10 million developers, and 45% of all new code written internally at Baidu is now generated by AI.

Under the hood, PaddlePaddle 3.2 debuted with an MFU (Model Flops Utilization) of 47% during pre-training runs on a massive 300B-parameter ERNIE model. Baidu stated that its PaddlePaddle-ERNIE ecosystem now counts 23.33 million developers and 760,000 enterprise users.

Why it matters

We think Baidu is executing a smart two-track playbook: deploying a closed flagship for high-margin cloud revenue while feeding open-source developers lightweight reasoning models.

For independent developers and resource-strapped engineering teams, ERNIE-4.5-21B-A3B-Thinking is the real standout. An active parameter count of just 3 billion means you can potentially self-host sophisticated chain-of-thought reasoning without spending a fortune on enterprise GPU clusters.

For the AI landscape, Baidu's decision to benchmark ERNIE X1.1 against DeepSeek R1-0528 proves that the reasoning race ignited by recent open breakthroughs has become the primary arena. Everyone has to prove they can match frontier reasoning while slashing hallucination rates.

However, Baidu's key hurdle remains global developer adoption. While 23 million developers on PaddlePaddle is massive domestically, winning mindshare outside China against the entrenched PyTorch and Hugging Face ecosystems requires consistent open-source credibility over time.