Deepseek OCR
Deepseek OCR is an open-weights language model from DeepSeek, released October 21, 2025. It has 3.3B parameters (mixture-of-experts, 1.1B active). The context window is 8K tokens.
We present DeepSeek-OCR as an initial investigation into the feasibility of compressing long contexts via optical 2D mapping. DeepSeek-OCR consists of two components: DeepEncoder and DeepSeek3B-MoE-A570M as the decoder.
Check it against your own PC →
Can you run Deepseek OCR locally?
The smallest common GPU it runs on is 6 GB of VRAM (with 32 GB of RAM): Q6_K, fully on GPU, roughly 110 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~110 |
| 8 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~150 |
| 12 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~190 |
| 16 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~240 |
| 24 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~380 |
| 32 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 4.3 GB | – | ~530 |
Ready-made quantizations on Hugging Face: sabafallah/DeepSeek-OCR-GGUF · NexaAI/DeepSeek-OCR-GGUF · NexaAI/DeepSeek-OCR-GGUF-CUDA · larshiakzemil/unsloth_finetune · see-k/unsloth_finetune · forkjoin-ai/deepseek-ocr-gguf
Benchmark results
Epoch AI has not published results for Deepseek OCR yet. New models are usually evaluated within days to weeks.
Links
Announcement · Hugging Face: deepseek-ai/DeepSeek-OCR
More from DeepSeek
DeepSeek V4.1 Flash 1017 · DeepSeek V4 Flash Vision Exp · DeepSeek V4.1 Flash · DeepSeek V4 Pro 0813 · DeepSeek V4 Flash 0731 · DeepSeek V4 Pro DSpark · DeepSeek V4 Flash DSpark · DeepSeek-V4-Pro