DeepSeek-OCR
by deepseek-ai
3.3B params · image-text-to-text · 3.3k likes · 2.3M downloads
DeepSeek-OCR is a 3.3B parameter model. At Q4 quantization it requires 2GB of VRAM. It runs comfortably on GeForce RTX 4090 (393 tok/s), GeForce RTX 5090 (590 tok/s), M4 Max 128GB (144 tok/s).
Inference providers
| Provider | $/1M in | $/1M out | Throughput |
|---|---|---|---|
| Novita | 203 tok/s |
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 393 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 590 tok/s | comfortable |
| M4 Max 128GB | 128GB | 144 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 72 tok/s | comfortable |
| M4 Pro 24GB | 24GB | 72 tok/s | comfortable |
| A100 PCIe 80 GB | 80GB | 721 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 1450 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 347 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 287 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 279 tok/s | comfortable |