What Happened
A developer has released EdgeDox, an Android app that runs document AI entirely on-device using Qwen3.5-0.8B, a lightweight language model from Alibaba's Qwen family. The app lets users ask questions about PDFs and other documents without sending any files to cloud servers.
The app is available on the Google Play Store and runs the 0.8 billion parameter model directly on the phone's processor. Current features include asking questions about PDF content, document summarization, and text extraction - all processed locally.
The motivation is straightforward: most document AI tools (Adobe Acrobat AI, AWS Textract, Azure Document Intelligence) require uploading files to external servers. For sensitive documents - legal contracts, medical records, financial statements - that is a non-starter for many users. EdgeDox eliminates that concern entirely by keeping everything on the device.
Why It Matters
On-device AI has been mostly theoretical for document processing. Apple and Google have added small models to their operating systems, but purpose-built document AI running locally on a phone is new territory.
The privacy implications are significant. Lawyers reviewing contracts, healthcare workers checking patient documents, financial advisors looking at statements - none of these people should be uploading sensitive files to third-party servers. Until now, the alternative was manual review or expensive enterprise solutions with compliance guarantees. A free app running everything locally removes that tradeoff entirely.
There is also an offline use case. Field workers, travelers, or anyone with unreliable connectivity can process documents without needing a data connection. This is the kind of capability that cloud-first tools structurally cannot provide.
The choice of Qwen3.5-0.8B is deliberate. At 0.8 billion parameters, it is small enough to run on mid-range Android hardware without draining the battery in minutes. The tradeoff is capability - do not expect GPT-4 level reasoning about complex legal language or nuanced financial analysis. But for straightforward questions like "what is the total amount due?" or "when does this contract expire?", a small model should handle it fine.
Our Take
This is an early proof-of-concept, not a replacement for serious document processing tools. A 0.8B parameter model will struggle with complex multi-page documents, tables with intricate formatting, or questions requiring cross-referencing across sections. The accuracy ceiling is real and matters for professional use.
But the direction is right. Most document AI questions are simple: "What is the effective date?", "Summarize section 3", "What are the payment terms?" These do not need 400 billion parameters and a data center. They need a competent small model running fast and privately.
The bigger signal here is that the gap between cloud AI and on-device AI is closing faster than expected. Qwen3.5-0.8B did not exist a year ago. In another year, we will likely have 2-3B parameter models running comfortably on phones with meaningfully better accuracy. Tools like Adobe Acrobat AI and Azure Document Intelligence still win on capability today, but their privacy tradeoff becomes harder to justify as local models improve.
If you handle sensitive documents on mobile and need quick answers without cloud exposure, EdgeDox is worth testing. Keep your expectations calibrated to what a sub-1B model can actually do, and you will find it useful for basic document queries.