DeepSeek-OCR

OCRdeepseek-ocr

OCR-focused multimodal stack for documents.

Use DeepSeek-OCR when you need resilient text extraction across mixed-quality captures. Same chat contract as your text models — swap the `model` field only.

Best for

Image OCR

Mode

OCR

Reasoning

No reasoning mode

Tools / coding

—

Throughput

860 tokens/sec

At a glance

  • Modalities: Image · Multimodal
  • Multimodal messages for scans
  • Works with guarded BFF patterns
  • Batch pages for throughput

Integration examples

These snippets are adapted for this model's API mode: OCR. The model field is set to deepseek-ocr. If your deployment uses an alias, mirror that value when you paste into your app. For general concepts, see the main documentation hub.

Deploy a model →

Keys and base URL

Replace YOUR_API_KEY with the key from Models after you sign in. This page's snippets use http://app.ai-grid.io:4000 when you are signed in (see DOCS_AUTHENTICATED_API_BASE_URL). Otherwise they keep the placeholder host for sovereign-safe documentation.

For PDFs, render each page to PNG/JPEG first, then send each page as image input. Keep files small enough for browser or gateway limits.

Extract text from an image

OCR

/v1/chat/completions
OCR · HTTP
IMAGE_DATA="$(base64 -w 0 document-page.png)"

curl http://app.ai-grid.io:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "deepseek-ocr",
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Extract all readable text. Preserve tables and line breaks." },
          {
            "type": "image_url",
            "image_url": {
              "url": "data:image/png;base64,'"$IMAGE_DATA"'"
            }
          }
        ]
      }
    ],
    "max_tokens": 1200
  }'

Other models

Open-weight · Text LLM

gpt-oss-120b

General agents

GO
TextMultimodalAdvanced reasoningToolsCode

gpt-oss-120b

Large open stack for general reasoning, code, and assistants.

Throughput

1,000 tokens/sec

  • Strong default for assistants and tool loops
  • OpenAI-compatible `/v1/chat/completions`
  • Pair with LangGraph for stateful agents
View documentation

Qwen · Text LLM

Qwen3-30B-A3B-Thinking

Reasoning

Q3
TextAdvanced reasoningToolsCode

Qwen3-30B-A3B-Thinking

Reasoning-forward 30B tier for planning and analysis.

Throughput

660 tokens/sec

  • Extended thinking style outputs
  • Ideal for LangGraph flows
  • Tune system prompts for chain-of-thought depth
View documentation

Google · Text LLM

google/gemma-4-31B

High-volume chat

G4
TextStandard reasoningToolsCode

google/gemma-4-31B

Fast 31B text generation for assistants and RAG answers.

Throughput

580 tokens/sec

  • Great for high-volume chat
  • Pairs with LangChain ChatOpenAI
  • Stable for LlamaIndex completion nodes
View documentation

Alibaba · Embedding

Alibaba-NLP/gte-Qwen2-7B-instruct

Embeddings

AB
TextNo reasoning mode

Alibaba-NLP/gte-Qwen2-7B-instruct

Dense text encoder for semantic search and clustering.

Throughput

No throughput info

  • Vector / similarity workflows
  • LlamaIndex Settings helper
  • Keep API id in sync with the model catalog
View documentation

Zhipu / Z.ai · OCR

zai-org/GLM-OCR

PDF OCR

GL
ImageMultimodalNo reasoning mode

zai-org/GLM-OCR

Document and image OCR with multimodal chat messages.

Throughput

1.96 pages PDF/sec

  • Vision + OCR style prompts
  • Great for ingestion pipelines
  • Resize images for latency
View documentation