Discussion on Lightweight Document Parsing Models and Benchmarks (DPT-2 mini / Gemini 3 Pro)

Landing AI has released DPT-2 mini, focusing on lightweight parsing and high throughput for digital native documents. There is also a discussion on the Gemini 3 Pro's status as SOTA in practical tool usage benchmarks.
Key insights
Positioning and Advantages of DPT-2 mini
The DPT-2 mini released by Landing AI is a lightweight model designed for high throughput and structured digital native documents. It emphasizes a small model size and fast parsing speed, making it suitable for fixed-format documents such as English digital PDFs, standard invoices, and digital contracts. It outputs structured Markdown key-value pairs and tables for subsequent processing.
Limitations and Use Cases of DPT-2 mini
The DPT-2 mini is explicitly not suitable for complex layouts, multilingual, handwritten, or scanned documents, which still require the more advanced DPT-2 model. It is recommended to call the model through Landing AI's ADE library, with a process that includes initializing the client, calling the parse method, and returning Markdown output, aimed at digital workflows such as financial reporting and contract review.
Performance of Gemini 3 Pro in Real Tool Usage Benchmarks
There are viewpoints indicating that the Gemini 3 Pro remains SOTA in real-world tool usage benchmarks (such as Vending-Bench) and several other benchmarks, providing comparative information that illustrates the competitive position of high-end multifunctional models in tool usage scenarios.