Baidu launches ERNIE 5.5 Flash and a marketplace for AI models

The new ERNIE 5.5 Flash Preview

Baidu recently unveiled ERNIE 5.5 Flash Preview, a new foundation model built for efficient inference. Inference is the process of actually running a model to generate an output, which can often be slow or expensive. This specific model aims to provide a balance between high-level intelligence and the speed required for practical use.

The model architecture includes 280 billion total parameters, though only 14 billion are activated per token. This design allows it to perform well in complex areas like 3D modeling, coding, and knowledge work while maintaining fast decoding speeds. Baidu plans to open-source the model weights by November 25, making it available to a wider range of developers.

To keep costs down, Baidu uses a proprietary technology called ERNIE Evolve. This is a recursive self-improvement method that helps the model become more efficient during its operation. The model is already available on Baidu AI Cloud’s Qianfan platform with specific pricing for input and output tokens.

Specialized models and agents

Alongside the foundation model, Baidu released two specialized application models. The ERNIE Wizshot Micro-Expression model is designed for high-quality video creation, specifically for podcasts and short animated dramas. It can process text, images, and speech simultaneously to create expressive content at a lower cost than typical industry standards.

The ERNIE Light Deep Thinking model handles complex requests by using a dual-model approach. It pairs a fast-thinking model that interacts with users with a slow-thinking model that refines the answers in a closed loop. It also uses Baidu Sparse Attention, a technology that compresses historical memory to save on costs without hurting the model's performance.

Baidu also updated its suite of AI agents to help with specific business tasks. These include Baidu Yijing, a digital human agent for video production, and DuMate, a general-purpose agent that serves as a gateway to other tools. The company also highlighted KOOKO, a platform for creative workflows that draws on billions of professional resources and academic papers.

A marketplace for AI services

To address the growing demand for diverse AI services, Baidu launched Token Exchange. This is a marketplace designed to be neutral and transparent, connecting model providers with customers. It allows users to compare different models side-by-side using clear data on pricing and performance metrics.

The platform relies on two main pillars to ensure quality: verification and credit ratings. Baidu verifies each resource by comparing model fingerprints and testing their capabilities. They also monitor calls in real time and perform regular spot checks to maintain a rating for every provider on the exchange.

In its first month, Token Exchange has already onboarded more than 100 model resources. As reported by Kosmo Online, the platform currently handles a daily average of more than 50 million calls. Baidu expects this marketplace to help put underused computing power to work and create a healthier ecosystem for foundation models.

Scaling infrastructure and hardware

Baidu is also strengthening the physical and cloud infrastructure that powers these models. The company launched the Tianchi 1024-card Supernode, which uses an all-optical interconnect for high-speed data transmission. This hardware is specifically built to support the demands of large-scale AI training and deployment.

On the hardware side, the Kunlunxin chip business announced a new roadmap. The Kunlunxin M100 inference chip has already shipped, while the flagship M300 is expected in 2027. Baidu's leadership believes that inference, rather than just training, will be the main driver for long-term growth in the AI chip market.

The company's autonomous driving division, Apollo Go, continues to expand its reach globally. It currently covers 28 cities and has logged over 350 million autonomous kilometers. This includes over 240 million kilometers driven in fully driverless mode, with commercial operations already running in Abu Dhabi and Dubai.

The vision for intelligent organizations

Baidu's CEO, Robin Li, framed these updates as part of a broader shift toward the 'intelligent organization.' He argues that companies must move away from traditional hierarchies to fully utilize AI. He suggests that organizations should be built around specific tasks rather than fixed job titles, creating a network of work instead of a ladder of roles.

Li identified three principles for this evolution: breaking down job boundaries, capturing tacit knowledge as a shared asset, and designing closed-loop processes. By turning practical experience into reusable skills, companies can build capabilities that benefit the entire workforce. He believes this creates a 'moat' that competitors cannot easily copy or hire away.

Baidu has already begun testing these principles internally. Over the last year, the number of daily active agents within the company rose from 5,000 to 37,000. These agents now complete more than 1 million tasks every week, demonstrating how the company intends to scale these AI tools for broader industrial use.