Model comparison
SmolVLM2-256M-Video-Instruct vs Qwen3-0.6B
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: SmolVLM2-256M-Video-Instruct's KV cache at 32K is 5.0× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| SmolVLM2-256M-Video-Instruct | Qwen3-0.6B | |
|---|---|---|
| Parameters | 256M | 752M |
| Architecture | llama | qwen3 |
| Layers | 30 | 28 |
| Native context | 8,192 | 40,960 |
| Mixture of experts | no | no |
| Quantizations published | 26 | 44 |
| Smallest quantization | 0.09 GiB | 0.20 GiB |
| Q4_K_M | — | 0.37 GiB |
| Licence | apache-2.0 | — |
KV cache by context
the term that decides long-context viability
| Context | SmolVLM2-256M-Video-Instruct | Qwen3-0.6B | Ratio |
|---|---|---|---|
| 4,096 | 0.09 GiB | 0.44 GiB | 4.98× |
| 8,192 | 0.18 GiB | 0.88 GiB | 4.98× |
| 16,384 | 0.35 GiB | 1.75 GiB | 4.98× |
| 32,768 | 0.70 GiB | 3.50 GiB | 4.98× |
| 65,536 | 1.41 GiB | 7.00 GiB | 4.98× |
| 131,072 | 2.81 GiB | 14.00 GiB | 4.98× |