DeepSeek-OCR-2: DeepSeek's 3.4B open OCR model that reads pages like a person
An open-weight OCR model that reads document images into Markdown or plain text at dynamic resolution, with a new DeepEncoder V2 that reorders visual tokens by the page's logical structure instead of scanning top-left to bottom-right.
Why it matters
Top of the OmniDocBench v1.5 leaderboard at 91.09% with the fewest visual tokens (256-1,120), beating Gemini-3 Pro on reading-order edit distance - in a 3.4B model that fits on a laptop.
What you could build with it
A law firm or research team could run DeepSeek-OCR-2 locally to convert scanned contracts and papers into searchable Markdown without sending sensitive documents to a cloud OCR service.
Does it hold up?
Usable now and genuinely small: runs on a laptop at 4-bit, with llama.cpp and vLLM support; but the documented setup pins old versions (CUDA 11.8, vLLM 0.8.5) and the model card gives no benchmark table, so verify accuracy on your own documents.
Built with DeepSeek-OCR-2
- DeepSeek-OCR 2: parameters, license and hardware needsai-opensource.com · The encoder change explained, parameter and memory tables, and the caveat that the model card publishes no benchmark table.
- DeepSeek OCR on a scanned PDF: Mac setup and results, testedbrasspage.com · A real test: 1,948 of 1,955 words read in order from a 10-page scan in 83 seconds on a Mac GPU via llama.cpp.
- DeepSeek Wins First Place Again: Pioneering 'Causal Flow' Visual Reasoning36kr.com · Coverage of the DeepEncoder V2 launch: 91.09% on OmniDocBench v1.5, better reading-order edit distance than Gemini-3 Pro, and lower repetition rates in production use.
Learn more
First spotted on hf: source.
More AI models
MiniMind: train your own 64M LLM from scratch in about 2 hoursA complete, from-scratch LLM training pipeline at 64M parameters - pretraining, SFT, LoRA, DPO, PPO/GRPO, tool use and…models · JEV 0.82VoiceStudio: the open-source, fully-local ElevenLabs alternativeA desktop app bundling 20 local speech-generation engines and 10 transcription options behind one UI — voice cloning…models · JEV 0.76Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73tiny-voca: a fully open 4.5M-parameter TTS trained for under $10 of computeA fully open 4.5M-parameter English TTS — code, weights and dataset under Apache 2.0 — trained from scratch in 7.7…models · JEV 0.67Qwen4-Max: Alibaba's new flagship model goes liveAlibaba's 2.4T-parameter flagship is now generally available via API, with a 1M-token context, native vision, and…models · JEV 0.67
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.