Deepseek-OCR: Nén quang học theo ngữ cảnh

Massively unexpected update from DeepSeek: a powerful, high-compression MoE OCR model.
> In production, DeepSeek-OCR can generate 33 million pages of data per day for LLMs/VLMs using 20 nodes (x8 A100-40G).
They want ALL the tokens. You're welcome to have some too. https://t.co/ks97gjFuhd pic.twitter.com/mXV08ifRle

— Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxesTex) October 20, 2025

DeepSeek-OCR has some weird architectural choices for the LLM decoder: DeepSeek3B-MoE-A570M
-> uses MHA, no MLA (not even GQA?)
-> 2 shared experts (like DeepSeek V2, but V3 only has 1)
-> quite low sparsity, activation ratio is 12.5%. For V3 it’s 3.52%, for V2 it’s 5%
-> not… pic.twitter.com/nOYptOn3OE

— elie (@eliebakouch) October 20, 2025

Letsss gooo! DeepSeek just released a 3B OCR model on Hugging Face 🔥

Optimised to be token efficient AND scale ~200K+ pages/day on A100-40G

Same arch as DeepSeek VL2

Use it with Transformers, vLLM and more 🤗https://t.co/n4kHihS3At

— Vaibhav (VB) Srivastav (@reach_vb) October 20, 2025

NEW DeepSeek OCR model that outperforms dots ocr while prefilling 3x less tokens pic.twitter.com/g9T93PndFb

— Casper Hansen (@casper_hansen_) October 20, 2025

🚀 DeepSeek-OCR — the new frontier of OCR from @deepseek_ai , exploring optical context compression for LLMs, is running blazingly fast on vLLM ⚡ (~2500 tokens/s on A100-40G) — powered by vllm==0.8.5 for day-0 model support.

🧠 Compresses visual contexts up to 20× while keeping… pic.twitter.com/bx3d7LnfaR

— vLLM (@vllm_project) October 20, 2025

🚨 DeepSeek just did something wild.

They built an OCR system that compresses long text into vision tokens literally turning paragraphs into pixels.

Their model, DeepSeek-OCR, achieves 97% decoding precision at 10× compression and still manages 60% accuracy even at 20×. That… pic.twitter.com/5ChoESanC8

— God of Prompt (@godofprompt) October 20, 2025

is it just me or is this deepseek paper really…weird? like the flagship results are all about compression ratios and they’re gesturing at implications for LLM memory but… it’s an OCR model? are they suggesting that LLMs should ingest OCR embeddings of screenshots of old notes?? pic.twitter.com/ptxkgANIeW

— will brown (@willccbb) October 20, 2025

DeepSeek-OCR: https://t.co/Hww4tubUiS

— Ray Fernando (@RayFernando1337) October 20, 2025

I quite like the new DeepSeek-OCR paper. It's a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn't matter.

The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language… https://t.co/AxRXBdoO0F

— Andrej Karpathy (@karpathy) October 20, 2025

Compress everything visually!

DeepSeek has just released DeepSeek-OCR, a state-of-the-art OCR model with 3B parameters.

Core idea: explore long-context compression via 2D optical mapping.

Architecture:

- DeepEncoder → compresses high-res inputs into few vision tokens;
-… pic.twitter.com/qbRTi8ViLY

— 机器之心 JIQIZHIXIN (@jiqizhixin) October 20, 2025