Cohere · 2026-08-27 · major
Cohere Parse — a 2.3B document model that turns PDFs into clean Markdown
Cohere Parse is a 2.3B vision language model that reads PDFs, slides and images and returns Markdown with tables, bounding boxes and image descriptions. It scores 79.2 on ParseBench at $1.50 per 1,000 pages and runs at 4.5 pages a second.
A small vision model built for document pipelines: Markdown out, tables kept, 4.5 pages a second, $1.50 per thousand pages.
Key specs
| Parse bench average | 79.2 |
|---|---|
| Throughput | 4.5 pages/sec |
Quick facts
| Maker | Cohere |
|---|---|
| Model | parse-v5.0 |
| Size | 2.3B parameters, about 4.6 GB |
| Input | PDF, PPT, base64 JPEG |
| Output | Markdown with HTML tables and bounding boxes |
| Price | $1.50 / 1,000 pages |
| Availability | Cohere API, Model Vault, Microsoft Foundry, AWS SageMaker |
Benchmarks
| GPT-5.5 | 84.4% | |
|---|---|---|
| Opus 4.8 | 84.3% | |
| Gemini 3.5 Flash | 81.8% | |
| Cohere Parse | 79.2% | |
| LlamaParse (Cost Effective) | 78.3% | |
| Mistral OCR 4 | 74.5% | |
| Databricks AI Parse | 72.4% | |
| Google Document AI | 57.3% | |
| AWS Textract | 53.3% |
Pricing
| Cohere API | $1.50 / 1,000 pages |
|---|---|
| Model Vault · vs the Cohere API at 50% GPU utilization | 23% less |
What is it?
Parse reads enterprise documents — claims, contracts, invoices, financial filings — and returns clean Markdown a retrieval pipeline can index directly. The model detects tables and embedded images rather than flattening them, and Cohere ships it as parse-v5.0 at 2.3 billion parameters, roughly 4.6 GB on disk.
How does it work?
Rather than a general chat model reading a page, this is a dedicated vision language model with an 8,192-token context, aimed at throughput. Output keeps document structure: lists stay lists, tables come back as HTML, images get descriptions, and every element carries bounding box coordinates so a downstream agent can point back at the page it came from. Nine major world languages are supported.
Why does it matter?
Document extraction is usually the slow, expensive first step of a RAG system. On Cohere's ParseBench comparison the model averages 79.2, ahead of AWS Textract at 53.3, Google Document AI at 57.3 and Databricks AI Parse at 72.4, while frontier chat models GPT-5.5 and Opus 4.8 still lead at 84.4 and 84.3. The trade is accuracy for cost and speed: $1.50 per 1,000 pages and 36 pages a second on an 8-GPU H100 node.
Who is it for?
teams building RAG and document pipelines
Frequently asked questions
- How much does Cohere Parse cost?
- Cohere Parse is priced at $1.50 per 1,000 pages through the Cohere API. Running it privately through Model Vault cuts inference cost about 23 percent against the API at 50 percent GPU utilization. Cohere's worked example puts a 13-million-page-per-month enterprise roughly $12,000 a month cheaper on Model Vault than at hyperscaler rates of $10 per 1,000 pages.
- Is Cohere Parse more accurate than GPT-5.5 for document parsing?
- No. On the ParseBench comparison Cohere published, GPT-5.5 averages 84.4 and Opus 4.8 averages 84.3, both ahead of Cohere Parse at 79.2. Parse's case is price and throughput instead: a 2.3B model at $1.50 per 1,000 pages and 4.5 pages a second, versus frontier chat models charged per token.
- Can Cohere Parse run on my own hardware?
- Yes. Alongside the Cohere API, Parse ships through Model Vault for single-tenant and on-premises deployment, plus Microsoft Foundry and AWS SageMaker. At about 4.6 GB the model is small enough to self-host, and Cohere reports 36 pages a second on an 8-GPU H100 node, or roughly 2,160 pages a minute.
- What file types and languages does Cohere Parse handle?
- Cohere Parse accepts PDF, PPT and base64-encoded JPEG files and covers nine major world languages. Output is Markdown containing document text, lists, tables rendered as HTML, bounding box coordinates and descriptions of embedded images, which is the shape a retrieval or agent pipeline can consume without extra cleanup.
Try it
Call the Cohere API with model parse-v5.0 on a PDF, PPT or base64 JPEG.