Why Is Baidu Rushing to Open‑Source ERNIE 4.5? – The New Power Anchor in an Open AI Landscape

At the end of June, Baidu announced the full open‑source release of its ERNIE 4.5 series, just weeks after DeepSeek publicly opened its R1 large model on January 20, 2025, an event that shook the AI community.

DeepSeek R1 quickly earned 75,000 GitHub stars, broke the monopoly of major AI firms, and within three months spawned hundreds of vertical‑domain models covering more than 20 industries. A top‑tier hospital’s CT‑image diagnostic system fine‑tuned on R1 reduced misdiagnosis by 27%; a Shenzhen quantitative hedge fund built a trading strategy on R1 that outperformed GPT‑4‑driven solutions by 11.3% annual return.

These data points challenged incumbents who previously claimed open‑source models would fall behind, prompting Baidu to frame its decision as a correction of past mistakes, likening it to OpenAI CEO’s admission of a historical error.

The author argues that open‑source dramatically lowers entry barriers, allowing enterprises and even individuals to deploy distilled versions of 1.5B‑14B models, expanding the AI market much like the early browser wars when Microsoft’s bundling of IE led to antitrust scrutiny.

The flood of DeepSeek‑related articles, livestreams, and news created a perception that DeepSeek was the best‑performing model, intensifying pressure on Baidu, iFlytek and other companies to act quickly.

Historically, AI giants relied on closed ecosystems to capture AI dividends, but open‑source disrupts this cycle. For example, an e‑commerce platform that fine‑tuned the open‑source R1 for its customer‑service chatbot saved ¥8 million annually in API fees and transformed accumulated dialogue data into a proprietary knowledge base.

This shift forces AI leaders to transition from “technology priests” to “ecosystem gardeners,” maintaining control while embracing openness. Future competition is likely to move toward data governance, domain‑specific fine‑tuning (often via reinforcement learning), cloud‑based collaborative intelligence, and tighter hardware‑software integration.

In a world where code and models flow freely, the author concludes that the new anchor of power is the user experience; neglecting users will inevitably erode market share.