Cohere Releases Parse 5, a 2.3-Billion-Parameter Vision-Language Model for Enterprise Document Extraction
Cohere's Parse 5 converts complex enterprise PDFs into structured Markdown, scoring 79.2 on its ParseBench evaluation against rivals like GPT-5.5 and AWS Textract.
Overview
Cohere has released Parse 5, model ID parse-v5.0, a 2.3-billion-parameter vision language model built to extract structured data from complex enterprise documents, according to InfoQ. The model launched on August 27, 2026, according to Cohere, and converts visually rich PDFs — including financial reports and scientific papers — into clean Markdown while providing bounding box coordinates for visual grounding, according to InfoQ.
What We Know
Architecture. Parse 5 is built on Cohere Labs’ open-weight North-Micro-Vision-Instruct architecture, combining a custom-trained 400 million-parameter native-resolution vision encoder initialized from SigLIP 2 SO400M with an in-house 2 billion-parameter language model based on Cohere’s Command A+ architecture, according to InfoQ. The vision encoder uses 2D Rotary Positional Embeddings and learned 1D positional embeddings to preserve document spatial structure, and the model has an 8,000-token context window, according to InfoQ. Cohere says Parse 5 processes documents and images across nine major world languages, according to Cohere.
Benchmark performance. Cohere evaluated Parse 5 using ParseBench, a benchmark InfoQ describes as comprising more than 2,000 human-verified enterprise pages spanning insurance, finance, and government sectors, according to InfoQ. Cohere’s own published results put Parse 5’s average ParseBench score at 79.2 — 87.0 on table extraction, 86.6 on content faithfulness, and 64.0 on semantic formatting — trailing frontier general-purpose models GPT-5.5 (84.4) and Opus 4.8 (84.3) but ahead of dedicated document-processing tools including AWS Textract (53.3) and Google Document AI (57.3), according to Cohere. InfoQ separately reported that Cohere states the model achieved “an average score of 79.2 across table extraction, content faithfulness, and semantic formatting,” according to InfoQ.
Throughput and pricing. Cohere says Parse 5 processes 4.5 pages per second, and that running the model through its single-tenant Model Vault deployment option can cut inference costs by up to 61% compared with standard API pricing at full hourly utilization, according to Cohere. Cohere illustrates the potential savings with an example: an enterprise accounts-payable workflow processing about 13 million document pages per month would save roughly $144,000 annually using Model Vault instead of the standard Cohere API, according to Cohere. API access to Parse 5 is priced at $1.50 per 1,000 pages, according to Cohere.
Availability. Parse 5 is generally available through the Cohere API, Model Vault, Microsoft Azure AI Foundry, and Amazon SageMaker on AWS, according to Cohere. Developers can also test the model through a free Hugging Face Space, and the underlying open-weight North-Micro-Vision-Instruct foundation model is separately available for local use, according to InfoQ. Cohere positions Parse 5 as an addition to its Compass search and retrieval stack, alongside its existing Embed and Rerank models, according to Cohere.
What We Don’t Know
Cohere has not disclosed adoption figures or named specific enterprise customers using Parse 5 since its release. Independent, third-party benchmark comparisons beyond Cohere’s own published ParseBench table have not yet been published.