Baidu App Makes Office Agent Free

🔑 Enhanced Key Takeaways

•Baidu's GUI-based interactive search components enable the Baidu App to process longer, more complex queries and integrate diverse content formats, including videos, graphics, and text, to provide richer results.
•The free offering of ERNIE Assistant task engine 2.0 aligns with Baidu CEO Robin Li's emphasis on 'Daily Active Agents' (DAA) as a key metric for AI success, focusing on how many agents actively complete tasks and create value in business processes.
•The upgrade to ERNIE Teacher likely incorporates advancements from recent ERNIE models, such as ERNIE 5.1, which was released in April 2026 and achieved leading performance on various benchmarks with significantly reduced pre-training costs.
•ERNIE Assistant task engine 2.0's top ranking on the PinchBench leaderboard signifies its strong capability in real-world agentic tasks, encompassing file operations, data summarization, web research, and creative content generation.
•Baidu's ERNIE 5.0 model, a foundation for these AI services, utilizes an ultra-sparse Mixture-of-Experts (MoE) architecture that integrates language, image, video, and audio within a single autoregressive framework for unified multimodal understanding and generation.

📊 Competitor Analysis▸ Show

Feature/Category	Baidu (ERNIE Assistant 2.0)	Alibaba (Qwen AI Assistant)	DeepSeek	OpenAI (GPT-4.5/5.x) / Anthropic (Claude Opus 4.8)
Pricing	Free for end-to-end office workflows	Paid membership tiers (starting ~$29.60/year), chatbot free, video credits separate	Low-cost reasoning models, some open-source	Paid API access, subscription models (e.g., Copilot Pro for Microsoft 365 at $30/month for Microsoft Copilot)
Office Agent Capabilities	Handles end-to-end office workflows (info gathering to delivery), multimodal (text, image, audio, video)	AI office assistant, chatbot, video generation credits	Reasoning models, multimodal understanding, mathematics, Chinese-language tasks	General AI assistant, integrated with ecosystems (e.g., Microsoft 365, Google Workspace) for tasks like email summarization, data analysis, document creation
Benchmarks (Agentic/Office Tasks)	Ranked first on PinchBench leaderboard. Baidu's GLM-5.2 scored 17.91 on OmegaUse-OfficeVal (human baseline 27.79), cheaper and faster than humans	Qwen3.7 Max scored 92.5% on PinchBench (August 2026 snapshot). Qwen3.7-Plus scored 17.51 on OmegaUse-OfficeVal, averaged $0.22 and 0.193 hours per task	DeepSeek-V4-Pro scored 14.48 on OmegaUse-OfficeVal. DeepSeek V4 Flash 0731 on PinchBench	Claude Opus 4.8-fast leads PinchBench at 93.5% (August 2026 snapshot). Nemotron 3 Ultra (NVIDIA) leads PinchBench at 90% (August 2026 snapshot). GPT-5.6 Luna, GPT-5.6 Sol also on PinchBench
Open-Source Strategy	ERNIE 4.5 planned open-source by June 2025. ERNIE 5.1 under Apache License	Qwen models are open-source, eliminating licensing costs for self-hosted deployments	Open-source and highly cost-efficient models	Varies by model and offering (e.g., some models are closed-source, others have open-weight versions)

🛠️ Technical Deep Dive

ERNIE 5.0 Architecture: Employs an ultra-sparse Mixture-of-Experts (MoE) architecture, integrating language, image, video, and audio within a single autoregressive framework. It uses a shared backbone for unified sequence modeling and modality-agnostic expert routing. The activation rate is below 3%, allowing for substantial capacity expansion without proportional computational overhead.
ERNIE 5.0 Training: Trained on a large, high-quality multimodal dataset, exposed simultaneously to text, images, videos, and audios from the beginning. Utilizes a unified Next-Group-of-Tokens Prediction objective for deep cross-modal interactions.
ERNIE 4.5-0.3B-Base Architecture: A dense transformer-based model with 18 layers, 1024 hidden dimensions, and 16 attention heads. It uses a Grouped-Query Attention (GQA) mechanism with 2 key-value heads, RMS Normalization, and the Swish (SiLU) activation function. Supports a context length of up to 131,072 tokens.
ERNIE-Image Architecture: Built upon a latent diffusion model (LDM) framework, utilizing an 8B single-stream Diffusion Transformer (DiT) architecture. It is paired with the Ministral-3 (3B) model as its text encoder and the FLUX.2 VAE for efficient latent space encoding and decoding.
General ERNIE Training Process: Typically involves pre-training on large datasets, followed by refinement using techniques such as supervised fine-tuning, reinforcement learning from human feedback, and prompt engineering.

🔮 Future ImplicationsAI analysis grounded in cited sources

The free availability of ERNIE Assistant task engine 2.0 will accelerate AI adoption in routine office tasks across China.

By removing cost barriers, Baidu encourages widespread experimentation and integration of AI into daily workflows, potentially setting a new industry standard for AI utility and accessibility.

Baidu's emphasis on 'Daily Active Agents' (DAA) will likely shift the industry's evaluation metrics from raw model performance to practical, value-creating AI applications.

This focus prioritizes the real-world utility and active deployment of AI agents in business processes, pushing developers to create more practical and integrated AI solutions that deliver tangible value.

The enhanced multimodal capabilities of ERNIE models will lead to more sophisticated and versatile AI assistants capable of handling diverse data types beyond text.

Integrating text, image, video, and audio processing within a single framework allows for more comprehensive understanding and generation, enabling AI to tackle complex, real-world scenarios that require processing multiple forms of information.

⏳ Timeline

2019

Baidu begins developing the ERNIE (Enhanced Representation through Knowledge Integration) series of large language models.

2023-03

Baidu launches ERNIE Bot for invited testing, based on ERNIE 3.0.

2023-08

ERNIE Bot is released to the general public after receiving regulatory approval in China.

2025-03

Baidu unveils ERNIE 4.5 and ERNIE X1 models, with ERNIE 4.5 featuring improved reasoning and multimodal capabilities.

2026-02

ERNIE 5.0, a 2.4 trillion-parameter unified multimodal foundation model with an ultra-sparse Mixture-of-Experts (MoE) architecture, is introduced.

2026-04

ERNIE 5.1 is officially released, topping multiple leaderboards and demonstrating comprehensive upgrades across agent, reasoning, and creative capabilities.

Baidu App Makes Office Agent Free

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (24)

👉Related Updates

Harness Turns AI Intelligence Into Production