Model comparison
DeepSeek-OCR-2 vs Qwen3-VL-8B-Instruct-abliterated-v1
At Q4_K_M, DeepSeek-OCR-2 is the smaller download — 1,950,326,688 bytes against 5,027,785,792.
From the file· summed bytes, KV per layer
Side by side
| DeepSeek-OCR-2 | Qwen3-VL-8B-Instruct-abliterated-v1 | |
|---|---|---|
| Parameters | 3.4B | 8.8B |
| Architecture | deepseek2-ocr | qwen3vl |
| Layers | 12 | — |
| Native context | 8,192 | — |
| Mixture of experts | yes, 64 experts | no |
| Quantizations published | 8 | 48 |
| Smallest quantization | 1.35 GiB | 1.97 GiB |
| Q4_K_M | 1.82 GiB | 4.68 GiB |
| Licence | — | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | DeepSeek-OCR-2 | Qwen3-VL-8B-Instruct-abliterated-v1 | Ratio |
|---|---|---|---|
| 4,096 | 0.12 GiB | — | — |
| 8,192 | 0.23 GiB | — | — |
| 16,384 | 0.47 GiB | — | — |
| 32,768 | 0.94 GiB | — | — |
| 65,536 | 1.88 GiB | — | — |
| 131,072 | 3.75 GiB | — | — |