Model comparison
SmolVLM-500M-Instruct vs embeddinggemma-300m
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: embeddinggemma-300m's KV cache at 32K is 8.6× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| SmolVLM-500M-Instruct | embeddinggemma-300m | |
|---|---|---|
| Parameters | 507M | 303M |
| Architecture | llama | gemma-embedding |
| Layers | 32 | 24 |
| Native context | 8,192 | 2,048 |
| Mixture of experts | no | no |
| Quantizations published | 2 | 10 |
| Smallest quantization | 0.41 GiB | 0.26 GiB |
| Q4_K_M | — | — |
| Licence | apache-2.0 | — |
KV cache by context
the term that decides long-context viability
| Context | SmolVLM-500M-Instruct | embeddinggemma-300m | Ratio |
|---|---|---|---|
| 4,096 | 0.16 GiB | 0.04 GiB | 4.44× |
| 8,192 | 0.31 GiB | 0.05 GiB | 6.15× |
| 16,384 | 0.63 GiB | 0.08 GiB | 7.62× |
| 32,768 | 1.25 GiB | 0.14 GiB | 8.65× |
| 65,536 | 2.50 GiB | 0.27 GiB | 9.28× |
| 131,072 | 5.00 GiB | 0.52 GiB | 9.62× |