Autor: Jakub Rusinowski · Ostatnia aktualizacja: 11 września 2026
Model library → DeepSeek-OCR → DeepSeek-OCR 3B
All 3B parameters stay resident — about 2.6 GB at Q4_K_M, so it runs on effectively any modern GPU — while the MoE decoder activates 6 of 64 routed experts plus 2 shared, roughly 570M parameters per token. Token budgets are explicit and chosen per page: Tiny 64 tokens at 512x512, Small 100 at 640x640, Base 256 at 1024x1024, Large 400 at 1280x1280, and a Gundam mode composing several 640x640 views with one 1024x1024 global view. 8,192-token sequence length. MIT licensed.
DeepSeek-OCR 3B needs about 3 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
| Parameters | 3 Billion (570M active) |
| Context window | 8,192 |
| Architecture | DeepEncoder + DeepSeek-3B MoE decoder |
| Provider | DeepSeek |
| Licence | MIT |
| Specified at | Q4_K_M |
| System RAM | 8 GB |
| Record updated | 2026-09-11 |
MIT — commercial use permitted. Commercial use permitted. No usage restrictions beyond attribution.
Modelled on a reference NVIDIA RTX 4090 (24 GB), with no KV cache (this record has no published architecture). Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.
| Quant | Weights | VRAM needed | Est. speed | Fit on 24 GB |
|---|---|---|---|---|
| Q2_K | 1.0 GB | 1.8 GB | ~328 tok/s (est.) | Fits comfortably |
| Q3_K_M | 1.3 GB | 2.1 GB | ~320 tok/s (est.) | Fits comfortably |
| Q4_K_M | 1.8 GB | 2.6 GB | ~307 tok/s (est.) | Fits comfortably |
| Q5_K_M | 2.1 GB | 2.9 GB | ~300 tok/s (est.) | Fits comfortably |
| Q6_K | 2.5 GB | 3.3 GB | ~292 tok/s (est.) | Fits comfortably |
| Q8_0 | 3.2 GB | 4.0 GB | ~278 tok/s (est.) | Fits comfortably |
| F16 | 6.0 GB | 6.8 GB | ~232 tok/s (est.) | Fits comfortably |
Want the memory numbers alone, at every quantization level and your own context length? Use the DeepSeek-OCR 3B VRAM calculator.
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
The cheapest catalogued GPU that runs DeepSeek-OCR 3B is the Intel Arc B570 (10 GB).
Install Ollama, then run:
ollama run deepseek-ocrWeights on Hugging Face: deepseek-ai/DeepSeek-OCR.
Best for: document analysis, vision, multimodal, edge devices.
← All DeepSeek-OCR models | VRAM calculator | Check your own hardware