Cohere Releases Parse, a 2.3B Multimodal Model for Enterprise Documents

Cohere Parse is now released

Cohere launched Parse on August 27, 2026, introducing a document-understanding model designed for high-volume enterprise parsing rather than general-purpose chat.

The model is identified in Cohere's release notes as parse-v5.0. Cohere says it is a 2.3-billion-parameter multimodal model with an 8K context window and an approximately 4.6 GB model size. It converts documents and images into structured output suitable for downstream AI systems, including Markdown or HTML content, HTML-formatted tables, page boundaries, visual-element locations and image descriptions.

Parse is available through Cohere's Parse API, Microsoft Foundry and AWS SageMaker. Organizations that need single-tenant deployment can also run it through Cohere Model Vault.

Beyond basic OCR

The product is aimed at documents where plain text extraction is not enough. Cohere says Parse recognizes reading order, tables, lists, forms, embedded images and captions while preserving document structure for retrieval and automation.

The company positions the model as a front end for enterprise search, retrieval-augmented generation and agentic workflows. Examples include extracting information from contracts, invoices and claims; preparing structured content for indexing; and giving multimodal agents better document context.

Cohere says the service supports nine major world languages and can be deployed in private-cloud or on-premises environments for organizations with tighter security and compliance requirements.

Pricing and benchmark claims

The hosted API is priced at $1.50 per 1,000 pages, according to Cohere. The company also publishes results on ParseBench, where it reports an average score of 79.2 across the three evaluation dimensions it uses for the product comparison.

Those benchmark numbers should be read as vendor-reported results. Cohere notes that its comparison excludes some ParseBench dimensions that are outside the current product scope. In particular, the current release does not target chart-series extraction, and Cohere says that capability is planned for a future Parser version.

Cohere's release notes provide a more concrete technical description than the marketing page: the model outputs structured document content and layout information but remains deliberately specialized rather than attempting every possible vision-language task.

Why a specialized parser matters

For enterprise AI, document ingestion is often a bottleneck. A strong retrieval or agent model cannot reason over information that has been lost during scanning, OCR or layout conversion. Tables, headings, captions and reading order can materially change meaning.

A smaller specialized model can therefore be valuable even when frontier general-purpose models score higher on some document benchmarks. The practical tradeoff is throughput, deployment footprint, predictable cost and control over where documents are processed.

Parse enters a crowded market that includes specialized OCR and document-intelligence services as well as multimodal frontier models. Its differentiation will depend on real-world parsing fidelity, latency, language coverage, deployment flexibility and total cost at production scale—not benchmark averages alone.

For developers, the key release status is clear: Parse is available now, rather than announced as an upcoming preview, and Cohere exposes it through both its own API and major enterprise AI platforms.