Cohere Releases Parse 5 Document Parsing Model

Parse 5 is a vision language model that takes PDFs, slides, and scanned images a page at a time and returns Markdown with the structure intact. Unlike OCR, which only reads characters, it recognizes tables, forms, diagrams, and embedded images, and it attaches bounding boxes, position coordinates, to tables and images so that a retrieval system or an agent can trace where in the original a piece of content came from. The company describes this output as the foundation for document indexing, RAG, and agentic retrieval. It supports nine languages, which according to reporting are Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish.

The first thing an engineer notices is the size. According to reporting, the model has 2.3 billion parameters, is built on Cohere Labs' North Micro Vision Instruct architecture, has an 8,192-token context window, and has a footprint of roughly 4.6 GB. A 400-million-parameter vision encoder reads the page at native resolution and holds its spatial structure with 2D rotary positional embeddings, and a 2-billion-parameter language model takes those features and writes Markdown in reading order. Because a page image goes in and Markdown comes out in a single pass through the model, the two-step pipeline of running OCR and then cleaning up with a separate model collapses into one. METAL reported on the release of North Micro Vision, the base model, as a 2.4-billion-parameter vision model.

The scores are in a table in the announcement. On the three-dimension average of ParseBench, a document-parsing benchmark, Parse 5 scores 79.2: 87.0 on tables, 86.6 on content faithfulness, and 64.0 on semantic formatting. In the same table, GPT-5.5 at 84.4, Opus 4.8 at 84.3, and Gemini 3.5 Flash at 81.8 sit above Parse 5, and the company notes that all three are general-purpose frontier models far larger than Parse. Below it are LlamaParse's Cost Effective tier at 78.3, the open-weight Chandra OCR 2 at 77.7, Mistral OCR 4 at 74.5, Databricks AI Parse at 72.4, and Azure Document Intelligence at 69.3, with Google Document AI at 57.3 and AWS Textract at 53.3. The company stresses that the gap to the two hyperscaler services is more than 20 points.

There are also cells missing from the table. According to the announcement's footnotes, the Layout and Chart dimensions of ParseBench were excluded from the comparison: Layout measures the ability to draw a bounding box around every text element, and this model is designed to attach coordinates only to tables and images; Chart measures the ability to pull numerical data out of a graph, and this model treats charts as visual elements with a descriptive caption. The company says chart data extraction is planned for the next version. The same footnote says the scoring rules as of August 2026 include a fix to bold and heading detection that had previously inflated semantic formatting scores, and that all competitor models were re-scored under the new rules.

According to reporting, Nils Reimers, VP of AI Search at Cohere, said in an interview, "Document parsing isn't solved because the hard part isn't reading text, it's preserving structure and meaning." Enterprise documents mix tables, diagrams, charts, and formatting that change how the data is interpreted, he explained, and most tools either drop structure or hallucinate content, while even frontier models break on layout-heavy pages. On why the model does not extract charts as data, he said, "Other solutions try to extract the data from the chart, but then miss out critical information (for example, the color or the pattern of a line) that leads to hallucinations in Chat and Agentic AI applications."

Speed and cost are the main event of this launch. According to the announcement, Parse 5 processes 4.5 pages per second on one GPU, and on an eight-H100 node that is 36 pages per second, or 2,160 pages a minute. On the same GPU configuration it is 1.4 times faster than RedNote's dots.mocr and 2.2 times faster than Chandra OCR 2, and the comparison was made only among open-source models, all served with vLLM. It can be used through the Cohere API or deployed on Model Vault, the company's single-tenant managed inference platform, and the company says Model Vault reduces inference cost by 23 percent compared with the API at 50 percent GPU utilization, with the saving growing to 61 percent at full utilization. The example scale the company gives is a large enterprise accounts-payable workflow processing 13 million pages a month.

According to reporting, Reimers said that when the company ran the numbers for a large financial services workflow processing 750 million documents a year, choosing Parse 5 over a large general-purpose model like GPT-5.5 would cut costs by more than 98 percent. That figure is Cohere's own estimate for a single modeled workflow, not an audited deployment. By analogy, having a frontier model read millions of invoices is like asking an accountant to sort the mail, and Parse 5 is a proposal to put a dedicated mail sorter in that seat. The fact that the accountant reads better stays right there in the table.

Outside observers look past the benchmark score to the next step. According to reporting, Stephanie Walter, Practice Leader for AI Stack at HyperFRAME Research, said, "It does not need to win every benchmark. It needs to make reliable enterprise-scale parsing economical," and pointed out that parsing is the first quality gate in the enterprise AI stack, so if tables, headings, images, or reading order are lost at ingestion, better embeddings and larger models cannot recover that structure. In the same report, Kevin Petrie, VP of Research at BARC US, said a survey his firm is completing shows document analysis is the number one AI use case, with 62 percent of the organizations polled having adopted it.

The announcement METAL reviewed also cites the makeup of ParseBench, and according to reporting the benchmark consists of more than 2,000 human-checked pages from insurance, finance, and government documents. The same report noted that on the full ParseBench leaderboard, LlamaParse's Agentic Plus configuration leads at 90.20, while what appears in Cohere's table is that company's cost-effective configuration. That is why, when reading the announcement's table, it matters which tier of each company's product was entered.

There are four deployment paths. It is available immediately through the Cohere API, Model Vault, Microsoft Foundry, and AWS SageMaker, and inside the Compass search stack it is bundled with the embedding model Embed and the reranking model Rerank so that ingestion, parsing, chunking, indexing, and hybrid search run through one interface. Teams in regulated industries can run it on their own infrastructure, and the announcement invites readers to try it on their own documents first in a free Space. METAL reported that Cohere CEO Aidan Gomez published an essay on who gets to set the rules for AI, and the company's emphasis on private-cloud and on-premises deployment carries straight through to how this product ships.

This is a launch in which the company published the rows where it loses on the accuracy table and set a throughput and cost table beside them. Few organizations can afford to hand every single page to a frontier model, and in workflows that handle millions of pages the contest is decided not by a five-point score gap but by cost per page and pages per second. Cohere wrote that calculation into its announcement, and the five-second clip on September 14 is that calculation compressed into a single line of advertising.