Cohere Parse 5: 2.3B Model Converts Enterprise Docs to Markdown at $1.50/1K Pages

Cohere has made parse-v5.0 generally available — no waitlist, no research license — through the Cohere Parse API, Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault. Built on Cohere Labs' North-Micro-Vision-Instruct architecture, the model carries 2.3 billion parameters, an 8,192-token context window, and a ~4.6 GB footprint. It accepts PDF, PPT, and JPEG pages as base64-encoded data URIs and returns structured Markdown in a single forward pass, with no separate OCR stage preceding it.

For engineers running RAG ingestion, claims pipelines, or contract search, the practical argument is straightforward: one model call replaces the traditional stack of OCR engine, layout detector, and table extractor. That kind of pipeline-level consolidation rather than raw model improvement is where most enterprise document processing gains are being found in 2026.

Architecture and Output

Parse 5 recovers text in human reading order (not raw PDF stream order), reconstructs tables as HTML to preserve merged cells and header rows, extracts form field label-to-value pairs, and emits image descriptions with bounding box coordinates — all in one pass. Nine languages have stable support: Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish. Zero-shot coverage extends further at reduced accuracy.

Two output modes are exposed. The default returns a per-page Markdown string ready for chunking and embedding. Setting output_format="blocks" returns typed block objects: a table block carries its rendered HTML, its bounding box, and a natural-language description. That second mode is what makes citation-level traceability viable — a retrieval system can map an answer back to exact page coordinates. What the model does not produce: per-element confidence scores, header or footer labels, font-hierarchy metadata, or JSON schema extraction. Pipelines that route low-confidence pages to human review must supply that routing logic externally.

Benchmark Claims and Their Limits

Cohere reports a ParseBench score of 79.2, placing Parse 5 ahead of Mistral OCR 4 (74.5), Azure Document Intelligence (74.3), and Databricks AI Parse (72.4) on that metric. ParseBench is a LlamaIndex benchmark covering approximately 2,078 human-verified enterprise pages across five scoring dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding.

The critical caveat: Cohere's 79.2 averages only three of those five dimensions and excludes charts and visual grounding — the two dimensions where most parsers lose the most ground. The public ParseBench leaderboard, which scores all five dimensions, tells a different story. On the full leaderboard, Mistral OCR 4 and Databricks AI Parse both sit at 60.68, and Azure Document Intelligence (Layout) at 59.64. Cohere Parse 5 does not currently appear on that leaderboard at all; LlamaParse Agentic leads it at 84.88.

Azure's three-dimension average works out to 74.3, which matches Cohere's reported figure exactly and validates the methodology — confirming that 79.2 is a vendor-reported subset score, not a leaderboard position. Engineers evaluating Parse 5 should treat it as a directional claim to reproduce on representative samples of their own documents, particularly those heavy in charts and visual-grounding content. For context on how OCR benchmarking methodology can shift competitive rankings, see our coverage of the Chandra OCR 2 open-source benchmark.

Pricing and Deployment

The metered API is cheaper at any volume below roughly 1.67M pages per month for a Medium instance and 2.87M pages per month for XL. Above those thresholds, dedicated capacity wins on unit economics — before accounting for data residency or air-gap obligations, which are typically the primary enterprise motivation for Model Vault regardless of cost. Cohere explicitly targets financial services, insurance, healthcare and life sciences, public sector, telecom, energy, and manufacturing — verticals where scanned forms and dense tables are the norm, and where residency constraints often make the API-versus-Vault decision before volume does.

For teams already operating agentic document interaction at scale, the $1.50-per-1,000-pages figure establishes a concrete unit cost that folds directly into pipeline cost models alongside embedding and inference spend.

Parse 5's release reflects a broader consolidation of the document-parsing layer around vision-language models that eliminate multi-stage preprocessing. At 2.3B parameters and ~4.6 GB, the model is small enough for constrained environments while producing output structures — HTML tables, bounding boxes, typed blocks — that align with what retrieval and agent pipelines consume downstream. The unresolved question is full-dimension benchmark parity: until Cohere submits to the public ParseBench leaderboard on all five dimensions, the 79.2 figure remains a directional claim, not a settled competitive position.