🐳DeepSeek-OCR-2: How to Run & Fine-tune

Guide on how to run and fine-tune DeepSeek-OCR-2 locally.

DeepSeek-OCR 2 is the new 3B-parameter model for SOTA vision and document understanding released on Jan 27, 2026 by DeepSeek. The model focuses on image-to-text with stronger visual reasoning, not just text extraction.

DeepSeek-OCR 2 introduces DeepEncoder V2, which enables the model to 'see' an image in the same logcal order as a human.

Unlike traditional vision LLMs that scan images in a fixed grid (top-left → bottom-right), DeepEncoder V2 builds a global understanding first, then learns a human-like reading order—what to attend to first, next, and so on. This boosts OCR on complex layouts by better following columns, linking labels to values, reading tables coherently, and handling mixed text + structure.

You can now fine-tune DeepSeek-OCR 2 in Unsloth via our free fine-tuning notebook. We demonstrated a 88.6% improvement for language understanding.

Running DeepSeek-OCR 2Fine-tuning DeepSeek-OCR 2

🖥️ Running DeepSeek-OCR 2

To run the model in vLLM, transformers or Unsloth, here are the recommended settings:

⚙️ Recommended Settings

DeepSeek recommends these settings:

• Temperature = 0.0

• max_tokens = 8192

• ngram_size = 30

• window_size = 90

Support Modes - Dynamic resolution:

• Default: (0-6)×768×768 + 1×1024×1024 — (0-6)×144 + 256 visual tokens

Prompts examples:

# document: <image>\n<|grounding|>Convert the document to markdown.
# other image: <image>\n<|grounding|>OCR this image.
# without layouts: <image>\nFree OCR.
# figures in document: <image>\nParse the figure.
# general: <image>\nDescribe this image in detail.
# rec: <image>\nLocate <|ref|>xxxx<|/ref|> in the image.

Turns any document into markdown using Visual Causal Flow.

📖 vLLM: Run DeepSeek-OCR 2 Tutorial

1. Obtain the latest vLLM via:

uv venv
source .venv/bin/activate
# Until v0.11.1 release, you need to install vLLM from nightly build
uv pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly

1. Then run the following code:

from vllm import LLM, SamplingParams
from vllm.model_executor.models.deepseek_ocr import NGramPerReqLogitsProcessor
from PIL import Image

# Create model instance
llm = LLM(
    model="unsloth/DeepSeek-OCR-2",
    enable_prefix_caching=False,
    mm_processor_cache_gb=0,
    logits_processors=[NGramPerReqLogitsProcessor]
)

# Prepare batched input with your image file
image_1 = Image.open("path/to/your/image_1.png").convert("RGB")
image_2 = Image.open("path/to/your/image_2.png").convert("RGB")
prompt = "<image>\nFree OCR."

model_input = [
    {
        "prompt": prompt,
        "multi_modal_data": {"image": image_1}
    },
    {
        "prompt": prompt,
        "multi_modal_data": {"image": image_2}
    }
]

sampling_param = SamplingParams(
    temperature=0.0,
    max_tokens=8192,
    # ngram logit processor args
    extra_args=dict(
        ngram_size=30,
        window_size=90,
        whitelist_token_ids={128821, 128822},  # whitelist: <td>, </td>
    ),
    skip_special_tokens=False,
)
# Generate output
model_outputs = llm.generate(model_input, sampling_param)

# Print output
for output in model_outputs:
    print(output.outputs[0].text)

🤗Transformers: Run DeepSeek-OCR 2 Tutorial

Inference using Huggingface transformers on NVIDIA GPUs. Requirements tested on python 3.12.9 + CUDA11.8:

torch==2.6.0
transformers==4.46.3
tokenizers==0.20.3
einops
addict 
easydict
pip install flash-attn==2.7.3 --no-build-isolation

from transformers import AutoModel, AutoTokenizer
import torch
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
model_name = 'deepseek-ai/DeepSeek-OCR-2'

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, _attn_implementation='flash_attention_2', trust_remote_code=True, use_safetensors=True)
model = model.eval().cuda().to(torch.bfloat16)

# prompt = "<image>\nFree OCR. "
prompt = "<image>\n<|grounding|>Convert the document to markdown. "
image_file = 'your_image.jpg'
output_path = 'your/output/dir'

res = model.infer(tokenizer, prompt=prompt, image_file=image_file, output_path = output_path, base_size = 1024, image_size = 768, crop_mode=True, save_results = True)

🦥 Unsloth: Run DeepSeek-OCR 2 Tutorial

1. Obtain the latest unsloth via pip install --upgrade unsloth . If you already have Unsloth, update it via pip install --upgrade --force-reinstall --no-deps --no-cache-dir unsloth unsloth_zoo

2. Then use the code below to run DeepSeek-OCR 2:

from unsloth import FastVisionModel
import torch
from transformers import AutoModel
import os
os.environ["UNSLOTH_WARN_UNINITIALIZED"] = '0'

from huggingface_hub import snapshot_download
snapshot_download("unsloth/DeepSeek-OCR-2", local_dir = "deepseek_ocr")
model, tokenizer = FastVisionModel.from_pretrained(
    "./deepseek_ocr",
    load_in_4bit = False, # Use 4bit to reduce memory use. False for 16bit LoRA.
    auto_model = AutoModel,
    trust_remote_code = True,
    unsloth_force_compile = True,
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for long context
)

prompt = "<image>\nFree OCR. "
image_file = 'your_image.jpg'
output_path = 'your/output/dir'
res = model.infer(tokenizer, prompt=prompt, image_file=image_file, output_path = output_path, base_size = 1024, image_size = 640, crop_mode=True, save_results = True, test_compress = False)

🦥 Fine-tuning DeepSeek-OCR 2

Unsloth now supports fine-tuning of DeepSeek-OCR 2. Like the original first model, you'll need to use custom semantic for it to work on transformers. Like the first model, Unsloth trains DeepSeek-OCR-2 1.4x faster with 40% less VRAM and 5x longer context lengths with no accuracy degradation. You can now fine-tune DeepSeek-OCR 2 via our free Colab notebook.

• DeepSeek-OCR 2: Fine-tuning only notebook

See below for CER (Character Error Rate) accuracy improvements on the Persian language:

Per-sample CER (10 samples)

idx	OCR1 before	OCR1 after	OCR2 before	OCR2 after
1520	1.0000	0.8000	10.4000	1.0000
1521	0.0000	0.0000	2.6809	0.0213
1522	2.0833	0.5833	4.4167	1.0000
1523	0.2258	0.0645	0.8710	0.0968
1524	0.0882	0.1176	2.7647	0.0882
1525	0.1111	0.1111	0.9444	0.2222
1526	2.8571	0.8571	4.2857	0.7143
1527	3.5000	1.5000	13.2500	1.0000
1528	2.7500	1.5000	1.0000	1.0000
1529	2.2500	0.8750	1.2500	0.8750

Average CER (10 samples)

• OCR1: before 1.4866, after 0.6409 (-57%)

• OCR2: before 4.1863, after 0.6018 (-86%)

📊 Benchmarks

Benchmarks for DeepSeek-OCR 2 model are derived from the official research paper.

Table 1: Comprehensive evaluation of document reading on OmniDocBench v1.5. V-token𝑚𝑎𝑥 represents the maximum number of visual tokens used per page in this benchmark. R-order denotes reading order. Except for DeepSeek OCR and DeepSeek OCR 2, all other model results in this table are sourced from the OmniDocBench repository.

Table 2: Edit Distances for different categories of document-elements in OmniDocBench v1.5. V-token𝑚𝑎𝑥 denotes the lowest maximum number of visual tokens.

Outperforms Gemini-3 Pro on the OmniDocBench