DeepSeek-OCR

Checkpointdeepseek-ai/DeepSeek-OCRModel card indexedModel card

Support summary

DeepSeek-OCR is a checkpoint published by DeepSeek, built on the DeepseekOCRForCausalLM architecture. vLLM 0.28.0, the newest indexed vLLM release, documents support. Official recipes or cookbooks cover SGLang, without pinning a stable release.

vLLM documents support in 20 of 20 indexed releases, from 0.17.0 to 0.28.0.

Documented hardware: AMD / ROCm and Intel GPU / XPU
Documented quantization: BF16
Configured context: 8,192 tokens
License: mit

Hardware, quantization and features count only statements about this checkpoint; engine-wide capabilities are listed with each release.

Hugging Face model card

Publisher: DeepSeek
Task: image text to text
License: mit
Library: transformers
Architecture: DeepseekOCRForCausalLM
Weight tensor elements: 3.3B
Tensor types: BF16
Configured context: 8,192 tokens
Languages: multilingual
Weight access: Public
Repository created: 2025-10-17

Source 1 Source 2 Source 3Reviewed 2026-09-05

Model notes and sources

Engine support

3 engines

Engine support for DeepSeek-OCR
Engine	Release	Evidence scope	Last known supported
vLLM	`0.28.0`Newest indexed	Checkpoint evidenceSource: vLLM v0.28.0 · docs/models/supported_models.mdReviewed 2026-09-05	`0.28.0`This release
Features, dependencies and evidence for vLLM 0.28.0
SGLang	Version not pinnedRecipe / cookbook	Documented recipe Official recipe Source 1: sglang · docs/cookbook/autoregressive/DeepSeek/DeepSeek-OCR.mdx Source 2: sglang · docs/src/snippets/autoregressive/deepseek-ocr-deployment.jsxReviewed 2026-09-05	No supported release indexed
Features, dependencies and evidence for SGLang 0.5.19
TensorRT-LLM	`1.2.1`Newest indexed	No matching evidence curated	No supported release indexed
Features, dependencies and evidence for TensorRT-LLM 1.2.1

Engine support

Recipes and cookbooks