Deepseek Ocr And The Unreasonable Usefulness Of Compression

DeepSeek-OCR and the Unreasonable Usefulness of Compression Announcing our Series F . Learn more

AI engineering

DeepSeek-OCR and the Unreasonable Usefulness of Compression

Why DeepSeek-OCR's 10x more efficient compression unlocks faster, cheaper intelligence, and how to run it performantly on Baseten

Authors

Alex Ker

Dhruv Singal

Last updated October 23, 2025

Share

TL;DR Why DeepSeek-OCR&#x27;s 10x compression unlocks faster, cheaper intelligence, and how to run it performantly on Baseten in \nConvert the document to markdown. 14 model_name: DeepSeek OCR 15 model_name: deepseek-ocr-latest 16 resources: 17 accelerator: H100_40GB 18 cpu: "2" 19 memory: 16Gi 20 use_gpu: true The next step is completing the model.py which includes the load and predict . Since we clone the original DeepSeek-OCR in our build commands, we can import some functions from the run_dpsk_ocr_image.py . We then construct load to match the DeepSeek image processor and initialize their async engine and predict to preprocess the image and pass the result through the processor. The last step is simply trussing the model with truss push --publish in the truss CLI and you should have the deployment visible in your Baseten dashboard. In the example repo, I’ve also included test_document_ocr.py to hit the deployed endpoint and visualizer.py to visualize (draw bounding boxes and label text) the OCR output mirroring the official implementation. You can see the returned result of the inference call in under a second from Baseten. ✕ Result of DeepSeek-OCR on Doctor’s note The takeaway The shift from text tokens to visual tokens isn&#x27;t just a technical optimization—it&#x27;s a fundamental rethinking of how we feed information to AI systems. DeepSeek-OCR proves that highly compressed data representations could be useful in a number of ways. Whether you&#x27;re building RAG pipelines, fine-tuning models without labelled data, or building real-time AI agents, consider trying out DeepSeek-OCR. You can spin this up on Baseten in <10 minutes and start experimenting with our sample code here .

Subscribe to our newsletter Stay up to date on model performance, inference infrastructure, and more. ‌

Explore Baseten today Start deploying Talk to an engineer

Related posts View all AI engineering

AI engineering How to run GLM-5.2 in any harness

Alex Ker

AI engineering Rolling deployments for zero-downtime model updates

Archit Mishra 2 others

AI engineering Cost-efficient, high-performance TTS with Qwen3-TTS

Ian Carrasco 1 other

Popular models GLM 5.2

Kimi K2.7 Code

DeepSeek V4

GPT OSS 120B

Whisper Large V3

NVIDIA Nemotron 3 Ultra

Explore all

Popular models GLM 5.2

Kimi K2.7 Code

DeepSeek V4

GPT OSS 120B

Whisper Large V3

NVIDIA Nemotron 3 Ultra

Explore all