Model comparison
SmolVLM2-256M-Video-Instruct vs LFM2.5-230M
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: LFM2.5-230M's KV cache at 32K is 1.9× 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 | LFM2.5-230M | |
|---|---|---|
| Parameters | 256M | 230M |
| Architecture | llama | lfm2 |
| Layers | 30 | 14 |
| Native context | 8,192 | 128,000 |
| Mixture of experts | no | no |
| Quantizations published | 26 | 25 |
| Smallest quantization | 0.09 GiB | 0.10 GiB |
| Q4_K_M | — | 0.14 GiB |
| Licence | apache-2.0 | other |
KV cache by context
the term that decides long-context viability
| Context | SmolVLM2-256M-Video-Instruct | LFM2.5-230M | Ratio |
|---|---|---|---|
| 4,096 | 0.09 GiB | 0.05 GiB | 1.87× |
| 8,192 | 0.18 GiB | 0.09 GiB | 1.87× |
| 16,384 | 0.35 GiB | 0.19 GiB | 1.87× |
| 32,768 | 0.70 GiB | 0.38 GiB | 1.87× |
| 65,536 | 1.41 GiB | 0.75 GiB | 1.87× |
| 131,072 | 2.81 GiB | 1.50 GiB | 1.87× |