Cohere Parse v5.0 for Tables, Forms and RAG Document Parsing

Model profile

Performance characteristics

2/10 Reasoning

1/10 Coding

8/10 Speed

9/10 Cost efficiency

Specifications

Technical details

Model family Parse

Model type Multimodal

Context window 8K tokens

Release date 2026-08-27

Status Live; generally available

Knowledge cutoff notes

Cohere has not published a direct knowledge-cutoff date for Parse v5.0. The model is designed for document parsing and should not be treated as a general-purpose knowledge model with a conventional training-data cutoff.

Model notes

Canonical model ID is parse-v5.0. The model is a 2.3B-parameter multimodal document parser with an approximate 4.6 GB model size and proprietary north-micro-vision-instruct architecture. Cohere documentation describes PDF, presentation, and JPEG workflows, but the current v2 Parse API reference documents image_url inputs, including remote HTTP(S) image URLs or base64 data URIs, and states that PDF and file URLs are not yet supported directly. Output can be Markdown or ordered blocks; tables may be represented as HTML and images include descriptions and bounding boxes. Cohere announced API pricing of $1.50 per 1,000 pages. Model Vault separately lists Parse 5 at $4 per hour or $2,500 per month for Medium and $7 per hour or $4,300 per month for XL. Editorial scores are comparative estimates for document-parsing use cases, not vendor specifications.

Cost

Model pricing

Input $1.50 per 1,000 pages through the Cohere API

Model guide

Cohere Parse v5.0: Document Parsing for Tables, Forms and RAG

Cohere Parse v5.0 is a 2.3-billion-parameter multimodal document-intelligence model that converts document images into structured Markdown or ordered content blocks. It is designed for enterprise search, retrieval-augmented generation, intelligent document processing and agent workflows rather than general-purpose conversation.

What is Cohere Parse v5.0?

Cohere Parse v5.0 is a specialized vision-language model for document intelligence. Instead of answering open-ended questions like a general-purpose chat model, it analyzes document pages and converts their visual content into structured text that other systems can search, index, retrieve and process.

The model is provided by Cohere and belongs to the company's Parse family of document-processing models. Its primary role is to prepare documents for downstream applications such as enterprise search, retrieval-augmented generation (RAG), intelligent document processing and AI agents. RAG systems retrieve relevant passages from a document collection before giving them to a language model; Parse helps create those passages from visually complex source files.

Cohere describes Parse v5.0 as using a proprietary north-micro-vision-instruct architecture. The supplied specifications list approximately 2.3 billion parameters, an approximate model size of 4.6 GB and an 8,192-token context length. These are model specifications, not guarantees of extraction quality for every document type.

What Parse v5.0 extracts

Parse is intended to preserve the structure that ordinary text extraction often loses. Its documented capabilities include:

Text in reading order
Tables and lists
Forms and key-value pairs
Images and image captions or descriptions
Page boundaries
Locations of visual elements using bounding-box coordinates

The principal output formats are Markdown and ordered content blocks. Markdown output can include HTML-formatted tables, image references, descriptions and positional information. The blocks format represents regions such as text, tables and images in document order, which can be useful when an application needs to render or post-process each region separately.

This makes Parse more than a basic OCR tool. OCR generally focuses on recognizing characters, while Parse attempts to retain document organization and relationships between visual elements. For example, a table can remain a table-like structure rather than becoming a sequence of unrelated lines, and fields in a form can be represented as associated key-value content.

Inputs, outputs and supported languages

The current v2 Parse reference documents image URLs and base64-encoded image data as the supported document input forms. It states that PDF and file URLs are not yet supported directly by that endpoint. Although Cohere documentation discusses PDF, presentation and JPEG document workflows, teams using the current v2 endpoint may need to convert PDF or presentation pages into images before submitting them.

This distinction matters in production. A workflow that receives PDFs cannot necessarily pass those files directly to the current Parse endpoint without preprocessing. Converting pages to images also adds an operational step and may affect image resolution, file handling and processing time.

Cohere documents stable support for Arabic, English, French, German, Japanese, Korean, Italian, Portuguese and Spanish. Other languages may work in zero-shot use, but the supplied documentation warns that accuracy can be lower. Organizations processing multilingual archives should validate representative documents rather than assuming identical performance across languages.

Parse produces text-oriented structured output, not generated images, audio or video. Its multimodal capability is therefore primarily on the input and document-understanding side: it reads visual document pages and returns Markdown or content blocks.

Context and output limits

The listed context length is 8,192 tokens. A context window is the amount of input and related processing information the model can handle in one request, although document-image limits and endpoint-specific restrictions can also affect practical usage.

The supplied research does not specify a maximum output-token limit. Parse is not documented as a general-purpose JSON-generation model, and the research lists native structured JSON-schema output as unsupported. Applications that need JSON records should parse the returned Markdown or blocks and validate the resulting fields in a separate application layer.

Parse also does not provide confidence scores according to the supplied documentation. If an application must identify uncertain fields, route low-quality pages for human review or attach field-level confidence values, those controls need to be implemented outside the model or verified through current Cohere product documentation.

Pricing and availability

Cohere announced API pricing of $1.50 per 1,000 pages. This page-based pricing is especially relevant to batch ingestion because the main cost driver is the number of pages processed rather than a conventional input-token and output-token calculation.

Parse v5.0 is generally available through the Cohere API and is also offered through Microsoft Foundry, AWS SageMaker and Cohere Model Vault, according to the supplied research. Model Vault uses separate deployment pricing. The listed Parse 5 Medium tier costs $4 per hour or $2,500 per month, while the XL tier costs $7 per hour or $4,300 per month. These are deployment prices rather than the standard per-page API price, so they should not be combined into a single rate.

Actual costs depend on the deployment route, page volume and operational requirements. Teams should confirm current commercial terms, supported regions and endpoint limits before committing to a production architecture.

Main strengths and limitations

Parse v5.0's main strength is its focus. It is designed to transform visually structured documents into useful downstream context, including tables, forms and page-level layout information. That focus can reduce the amount of custom extraction logic required before documents enter a search or RAG pipeline.

Its page-based API pricing is another practical advantage for organizations processing large, predictable document volumes. The model's relatively small 2.3-billion-parameter size is also consistent with a specialized processing model rather than a large general-purpose conversational system. The comparative speed and cost scores in the supplied research are editorial estimates, not Cohere-published benchmarks; they rate Parse highly for speed and cost within document-parsing use cases.

Important limitations include the following:

The current v2 reference does not document direct PDF or file-URL input, so preprocessing may be required.
There is no documented confidence-score output.
It does not provide native JSON-schema output in the supplied specifications.
It is not a general-purpose chat, reasoning or coding model.
It does not generate images, audio or video.
It may not identify every presentation-level detail, including all headers, footers or font hierarchy.
Languages outside the documented stable set may produce less reliable results.

Charts and other visual elements can be located and described, but the research does not establish that Parse reliably converts every chart into accurate, structured numerical data. Teams with strict financial, legal or compliance requirements should add validation and human review where extraction errors could materially affect decisions.

Reasoning, coding and tool support

Parse v5.0 should not be evaluated as a reasoning model in the same way as a conversational language model. Its documented task is visual document interpretation and structured extraction. The supplied comparative reasoning score is 2 out of 10 and the coding score is 1 out of 10; these are editorial assessments for this model's intended use, not vendor specifications or benchmark results.

The model is not documented as supporting tool or function calling, streaming or code execution. It can provide valuable input to an agent, but the surrounding application would normally use another model or software components for planning, tool selection, database actions, calculations and final responses.

Best use cases for Parse v5.0

Enterprise RAG ingestion: Convert reports, forms and other visual documents into content that can be indexed and retrieved.
Search indexing: Preserve reading order, tables and page information when building search systems over document collections.
Intelligent document processing: Extract information from forms, structured pages and semi-structured business records.
Document-grounded agents: Supply an agent with cleaner document context before it performs a supported workflow.
High-volume batch processing: Use page-based API pricing for predictable document-ingestion workloads.

Parse can be paired with Cohere Embed and Rerank in a retrieval pipeline: Parse converts pages into searchable content, Embed can represent content for semantic retrieval, and Rerank can help order candidate results. Those sibling capabilities are complementary; they do not turn Parse itself into a chat or agent model.

When to choose this model

Choose Parse v5.0 when the central problem is extracting usable structure from document images at scale. It is a sensible option when tables, forms, reading order and page-level visual information matter more than open-ended conversation, coding or complex reasoning. Its page-based pricing may also be attractive when document volume is easy to estimate.

Consider another option when the workflow requires direct PDF handling through the current endpoint, native JSON-schema responses, field-level confidence scores, detailed typography analysis or reliable chart-data extraction. A general-purpose multimodal model may be more appropriate for interactive questions about a small number of documents, while a dedicated OCR or document-processing system may be preferable when the application needs tightly controlled field validation. A separate language model is still needed for generation, coding, planning or tool use.

Bottom line

Cohere Parse v5.0 is best understood as a document-ingestion component rather than an all-purpose AI assistant. Its value comes from turning visually complex pages into ordered Markdown or content blocks that downstream search, RAG and enterprise-processing systems can use. The strongest fit is high-volume, structure-sensitive extraction; the main trade-offs are endpoint input constraints, the absence of documented confidence and JSON-schema output, and the need for other components when an application requires reasoning, generation or actions.

Answers to Frequently Asked Questions

What are the main limitations of Cohere Parse v5.0?

Parse v5.0 does not provide documented confidence scores or native JSON-schema output, and the current endpoint may require PDF preprocessing. It is also not a general-purpose chat, reasoning, coding or tool-use model. Languages outside its documented stable set may be less accurate, and chart-to-data extraction may require additional validation.

How much does Cohere Parse v5.0 cost?

Cohere announced API pricing of $1.50 per 1,000 pages. Deployment through Cohere Model Vault uses separate pricing: the listed Parse 5 Medium tier costs $4 per hour or $2,500 per month, while the XL tier costs $7 per hour or $4,300 per month. Current prices, regions and endpoint limits should be confirmed with Cohere.

Does Cohere Parse v5.0 support PDF files directly?

The current v2 Parse reference documents image URLs and base64-encoded image data as supported inputs, but it does not document direct PDF or file-URL input. Applications receiving PDFs may need to convert each page into an image before submitting it.

What is Cohere Parse v5.0 used for?

Cohere Parse v5.0 is a vision-language document parsing model that converts visually complex pages into structured Markdown or ordered content blocks. It is designed for enterprise search, RAG ingestion, intelligent document processing and document-grounded AI agents.

Can Cohere Parse v5.0 extract tables and forms?

Yes. Parse v5.0 is designed to preserve tables, lists, forms, key-value pairs, reading order, images, page boundaries and bounding-box coordinates. Tables may be returned as HTML-formatted tables in Markdown, while forms can be represented as associated key-value content.