Model comparison
Llama-3.2-3B-bnb-4bit vs Qwen3-VL-8B-Instruct-abliterated-v1
At Q4_K_M, Llama-3.2-3B-bnb-4bit is the smaller download — 2,019,373,344 bytes against 5,027,785,792.
From the file· summed bytes, KV per layer
Side by side
| Llama-3.2-3B-bnb-4bit | Qwen3-VL-8B-Instruct-abliterated-v1 | |
|---|---|---|
| Parameters | 3.3B | 8.8B |
| Architecture | llama | qwen3vl |
| Layers | 28 | — |
| Native context | 131,072 | — |
| Mixture of experts | no | no |
| Quantizations published | 1 | 48 |
| Smallest quantization | 1.88 GiB | 1.97 GiB |
| Q4_K_M | 1.88 GiB | 4.68 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Llama-3.2-3B-bnb-4bit | Qwen3-VL-8B-Instruct-abliterated-v1 | Ratio |
|---|---|---|---|
| 4,096 | 0.44 GiB | — | — |
| 8,192 | 0.88 GiB | — | — |
| 16,384 | 1.75 GiB | — | — |
| 32,768 | 3.50 GiB | — | — |
| 65,536 | 7.00 GiB | — | — |
| 131,072 | 14.00 GiB | — | — |