junkim100

/

DeepSeek-3B-MoE-decoder

like 0

Thanks for the interest! Just to clarify, this repo only contains the decoder component extracted from DeepSeek-OCR, not a standalone OCR model. You'd need to pair it with a vision encoder to do actual OCR.

For performance benchmarks on the full model, check out the original repo: https://github.com/deepseek-ai/DeepSeek-OCR

I had originally planned to use this Moe as an LLM decoder, as part of a on-device multi-modal large model. However, I've now found that the general performance metrics of this LLM decoder have decreased, which is a bit disappointing.