Baidu Pledges Apache 2.0 for ERNIE 5.5 Flash Preview Weights

Baidu has pledged to release the ERNIE 5.5 Flash Preview model weights under the Apache 2.0 license by late November. This comes alongside the already available ERNIE 5.5 Flash Base, which uses a restrictive custom license excluding organizations in the EU…

Christopher Ort
By Christopher Ort · 3 min read
Share
Server racks in a modern data center illuminated by blue lights during AI model training
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. What happened
  3. Why it matters now
  4. Who is most affected
  5. About the analysis
  6. i10x Perspective

Baidu just dropped a surprise legal move in the AI race. It pledged to release weights for the upcoming ERNIE 5.5 Flash Preview under the permissive Apache 2.0 license later this month. That announcement, however, sits beside the already-released ERNIE 5.5 Flash Base model, which carries a custom license that bars organizations in the European Union, the United Kingdom, and South Korea from any use, including basic testing.

i10X Quick Take

Quick Take

  1. The split release: Baidu’s ERNIE 5.5 Flash Base is out on Hugging Face but geoblocked via license, while the ERNIE 5.5 Flash Preview remains API-only pending a promised November weight release.
  2. Geopolitical fencing: The Base model’s custom agreement explicitly prohibits use, modification, and deployment by organizations operating within the EU, the UK, and South Korea.
  3. The commercial ceiling: Organizations outside the banned regions can use the Base model commercially for free, but hit a strict prior-authorization wall once related revenues exceed US$50 million.
  4. Why it matters: It highlights a growing trend of “open-washing” where highly capable models are distributed globally on open repositories but laden with legal tripwires for enterprise compliance teams.

What happened

Baidu released the ERNIE 5.5 Flash Base model, a 280-billion-parameter Mixture-of-Experts system that activates 14 billion parameters per token and supports a 1-million-token context window. The weights sit on Hugging Face in FP8 and BF16 formats, yet a custom license blocks any organization in the EU, UK, or South Korea from touching the model, even for internal tests.

Days later the company adjusted its message. It promised that the fine-tuned ERNIE 5.5 Flash Preview weights would land under Apache 2.0 by late November. For now the Preview version stays API-only through Qianfan and OpenRouter, so the Apache commitment remains a promise until the files and license text actually appear.

Why it matters now

The split between the two licenses reveals how open-weight models are becoming tools of regional policy. By locking the EU and UK out of the Base model, Baidu avoids the heavy compliance load of the EU AI Act and similar rules. The result is a fractured landscape where access depends as much on geography as on technical merit.

The rollout also shows the gap between “open weights” and true open source. Developers can download files from Hugging Face, yet that does not grant legal clearance to run them. Rules on derivatives, distillation, and synthetic data, plus the US$50 million revenue trigger for prior approval, raise the compliance burden for teams that adopt non-Western base models. Should the Apache 2.0 release happen, it would create an odd outcome: a free, globally usable fine-tune resting on a heavily restricted foundation.

Who is most affected

Enterprise legal teams, open-source developers, and AI startups operating across borders feel the effects most. Teams outside China now need to track exactly which ERNIE 5.5 variant they integrate, because mixing Base and Preview components could carry restrictive terms into later projects.

StakeholderImpactWhy
EU & UK DevelopersTotal exclusion from Base model.The custom license explicitly forbids use, even for non-commercial testing or research.
Scaling AI StartupsAuthorization bottleneck.Hitting US$50 million in annual related revenue triggers a mandatory written Baidu approval process.
AI Tooling BuildersAmbiguous derivative limits.Using the Base model for distillation or synthetic data generation carries strict downstream restrictions.

About the analysis

Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.

This is an independent, research-based i10x analysis designed for CTOs, AI developers, and legal compliance teams navigating the complexities of global AI models.

i10x Perspective

Baidu’s split-license approach signals the next phase for foundational models: strategic assets shaped by regulatory and territorial goals. The earlier vision of borderless open-source AI is giving way to a patchwork where license terms follow policy lines. When intelligence functions as a core economic input, these geoblocked agreements act as control valves that decide which regions gain access and which must develop alternatives on their own.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

ModelsNews

Editor's picks

  1. Baidu Pledges Apache 2.0 for ERNIE 5.5 Flash Preview Weights
  2. Scotland AI Data Centers Face Mounting Protests Over Energy Demands
  3. OpenAI Launches ChatGPT Wallet for Agent Purchases
  4. LLM 0.37.2 fixes OpenAI API key leak to third-party endpoints
  5. CVE-2026-110199: IBM ContextForge MCP Gateway XSS flaw fixed in version 1.1.0