Release of Specialized/Vertical Models (DeepSeekMath-V2 / DPT-2 mini)

Two vertical models have been released, covering advanced mathematical reasoning and structured document parsing. • DeepSeek has launched DeepSeekMath-V2, which has performed excellently in IMO-ProofBench and assessments like IMO 2025 and Putnam 2024, achieving scores close to or reaching the IMO gold medal level of 118/120 in Putnam. • DeepSeekMath-V2 is open-sourced and available for download on GitHub and Hugging Face under the Apache-2.0 license. • Landing AI has released DPT-2 mini, optimized for digital-native documents (invoices, contracts, simple forms), featuring a lightweight design and high throughput, outputting structured Markdown. • DPT-2 mini is not suitable for complex layouts, multilingual, or handwritten scanned documents, and is recommended for batch extraction tasks with fixed templates.
Key insights
DeepSeekMath-V2 Achieves High-Level Results in Competitive Mathematical Reasoning and is Open-Sourced
DeepSeek has released DeepSeekMath-V2, which has shown outstanding performance in problems from IMO-ProofBench, IMO 2025, and Putnam 2024, with scores approaching 118/120, indicating it has reached the level of an International Mathematical Olympiad gold medal. This model is open-sourced under the Apache-2.0 license on GitHub and Hugging Face, facilitating research and reproducibility.
DPT-2 mini Designed for High Throughput Structured Digital Document Parsing
Landing AI's DPT-2 mini is a lightweight parsing model designed for digital-native documents, emphasizing speed and consistency. It is suitable for batch extraction of standardized invoices, contracts, and other fixed templates, producing structured Markdown results. This model is not suitable for complex layouts or handwritten/scanned documents, making it ideal for integration into digital workflows to reduce computational overhead.