Cohere Releases Parse 5 for Faster Document Data Extraction

Cohere has unveiled Parse 5 (parse-v5.0), a multimodal vision-language model built to streamline one of enterprise development's thorniest tasks: pulling structured information out of messy, visually complex documents like financial reports and scientific papers. Rolled out on August 27th 2026, the 2.3-billion-parameter system converts PDFs directly into clean Markdown while also returning bounding box coordinates that link extracted data back to specific locations in the source file.

How It Works#

Parse 5 couples a custom-built 400M-parameter vision encoder (starting from Google's SigLIP 2 architecture) with a 2B-parameter language model built on Cohere's Command A+ design. The vision encoder uses 2D rotary positional embeddings to capture document layout, then funnels visual features into the language model through a dedicated projector. Rather than stacking outputs in sequence, the system injects patch embeddings from multiple encoder layers into the early layers of the language model—a technique borrowed from the DeepStack integration method—so the reasoner has access to visual features at different levels of abstraction.

The result sidesteps traditional brittle OCR pipelines, producing structured Markdown directly and eliminating the need for post-processing rules.

Performance and Availability#

On ParseBench, a curated benchmark of over 2,000 human-verified enterprise pages drawn from insurance, finance, and government sectors, Parse 5 scored 79.2 across table extraction, content faithfulness, and semantic formatting—a solid showing that trails only premium competitors like LlamaParse Agentic Plus (90.20) but outpaces Mistral OCR (75.05) and Google Gemini 3 Flash Thinking High (75.05).

Developers can access the API via Cohere's platform, Microsoft Azure AI Foundry, and Amazon SageMaker. The company also offers a free Hugging Face Space for browser-based testing and released the underlying North-Micro-Vision-Instruct foundation model as open-weight code on Hugging Face for local experimentation and fine-tuning.

Why It Matters#

The bounding box output is particularly valuable for regulated industries—insurance, finance, legal—where audit trails and human verification are non-negotiable. Downstream applications can pinpoint exactly where in the original document a piece of extracted data came from, closing a loop that traditional OCR tools leave open.