DeepSeek's Latest Open-Source Model is Taking Silicon Valley by Storm
Want to read in a language you're more familiar with?
DeepSeek's newest open-source model is creating a major buzz. Its elegance lies in its simplicity: a compact 3B parameter model delivering performance that chal...
DeepSeek's newest open-source model is creating a major buzz. Its elegance lies in its simplicity: a compact 3B parameter model delivering performance that challenges larger models. Some even speculate it might have open-sourced techniques closely guarded by giants like Google Gemini.
A potential hurdle? Its somewhat misleading name: DeepSeek-OCR.
This model tackles the computational challenge of processing long text contexts. The core, revolutionary idea is using vision as a compression medium for text. Since an image can contain vast amounts of text while consuming fewer tokens, the team explored representing text with visual tokens—akin to how a skilled reader can grasp content by scanning a page rather than reading every word. A picture is worth a thousand words, indeed.
Their research confirmed that with a compression ratio under 10x, the model's OCR decoding accuracy hits an impressive 97%. Even at a 20x ratio, accuracy remains around 60%.
Demonstrating remarkable efficiency, their method can generate over 200,000 pages of high-quality LLM/VLM training data per day using just a single A100-40G GPU.
Unsurprisingly, the release quickly gained traction, amassing 3.3K GitHub stars and ranking high on Hugging Face trends. On X, Andrej Karpathy praised it, noting that "images are simply better LLM input than text." Others hailed it as "the JPEG moment for AI," opening new pathways for AI memory architecture.
Many see this unification of vision and language as a potential stepping stone toward AGI. The paper also intriguingly discusses AI memory and "forgetting" mechanisms, drawing an analogy to how human memory fades over time—potentially paving the way for infinite-context models. The Core Technology
The model is built on a "Contextual Optical Compression" framework, featuring two key components:
- DeepEncoder: Compresses high-resolution images into a small set of highly informative visual tokens.
- DeepSeek3B-MoE-A570M: A decoder that reconstructs the original text from these compressed tokens.
The innovative DeepEncoder uses a serial process: local feature extraction on high-res images, a 16x convolutional compression stage to drastically reduce token count, and finally, global understanding on the condensed tokens. This design allows it to dynamically adjust "compression strength" for different needs.
On the OmniDocBench benchmark, DeepSeek-OCR achieved new SOTA results, significantly outperforming predecessors while using far fewer visual tokens.
