Cohere released Parse, a 2.3B model that turns PDFs into Markdown

Cohere released Parse on 27 August, a 2.3 billion parameter vision language model built to turn visually complex PDFs into clean Markdown with bounding box coordinates. InfoQ reported on the release on 3 September, which brought it back to the attention of teams evaluating document ingestion this week.

What Parse is

Parse is a single model, not a pipeline of OCR plus layout plus table extraction. It has a 400 million parameter native-resolution vision encoder in front of a 2 billion parameter language core, an 8,000 token context window, and covers nine major world languages. It is built on Cohere Labs' open-weight North-Micro-Vision-Instruct architecture, and Cohere offers on-premises and private cloud deployment alongside the hosted API.

The output is Markdown with visual grounding, so a downstream RAG pipeline can trace a quoted passage back to the coordinates on the page it came from. Document search systems usually build this part by combining three separate components.

The benchmark, attributed

Cohere reports a 79.2 average on ParseBench, its own benchmark covering table extraction, content faithfulness and semantic formatting. On the same benchmark Cohere puts AWS Textract at 53.3, Google Document AI at 57.3 and Azure Document Intelligence at 69.3. The newest, most capable general-purpose models score higher: Cohere lists GPT-5.5 at 84.4, Opus 4.8 at 84.3 and Gemini 3.5 Flash at 81.8. InfoQ notes that LlamaParse Agentic Plus leads at 90.20 and that Mistral OCR scores 75.05, with Gemini 3 Flash at the same 75.05.

Read those in the right order. Parse is a small model priced for volume that beats the three big cloud providers' document services and scores close to the smaller general models, while the most capable LLMs and an agentic parser rank ahead of it. Every number here is Cohere's, on Cohere's benchmark: treat it as the vendor's marketing, not an independent ranking.

Price and throughput

The hosted API is $1.50 per 1,000 pages. Cohere reports 4.5 pages per second on a single GPU and 36 pages per second on an 8-way H100 node, and says a self-hosted Model Vault deployment saves 23 to 61 percent versus the API depending on utilisation. It quotes a worked example at 13 million pages a month, where Model Vault saves about $144,000 a year against the API and about $1.47 million against unnamed cloud provider services.

If you have a document extraction step and you are still running one of the three cloud provider services Cohere benchmarks against, the vendor's own numbers put Parse ahead of your current service on quality at a price that is easy to plan against. That is enough to justify an evaluation on your own documents, especially if you would rather deal with one model than an OCR plus layout plus tables pipeline. If you are already using one of the most capable LLMs for the same job, the question is different: whether the general model's higher score is worth its higher per-page cost on your volume.

For anything you decide to run in production, do not take the ParseBench numbers as neutral. Run the parser on a representative sample of the documents you actually have, keep the outputs, and compare them the same way against whatever you use today.

Source

Primary source: Introducing Parse: Enterprise document intelligence at scale, Cohere, 27 August 2026. Coverage: Cohere's Parse 5 Promises Efficient Multi-Modal Information Extraction From Complex Documents, InfoQ, 3 September 2026.