Model comparison
SmolLM2-135M-Instruct vs tinygemma3_cifar
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: tinygemma3_cifar's KV cache at 32K is 5.7× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| SmolLM2-135M-Instruct | tinygemma3_cifar | |
|---|---|---|
| Parameters | 135M | 39M |
| Architecture | llama | gemma3 |
| Layers | 30 | 8 |
| Native context | 8,192 | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 41 | 1 |
| Smallest quantization | 0.08 GiB | 0.04 GiB |
| Q4_K_M | 0.10 GiB | — |
| Licence | apache-2.0 | wtfpl |
KV cache by context
the term that decides long-context viability
| Context | SmolLM2-135M-Instruct | tinygemma3_cifar | Ratio |
|---|---|---|---|
| 4,096 | 0.09 GiB | 0.06 GiB | 1.41× |
| 8,192 | 0.18 GiB | 0.08 GiB | 2.28× |
| 16,384 | 0.35 GiB | 0.09 GiB | 3.79× |
| 32,768 | 0.70 GiB | 0.12 GiB | 5.67× |
| 65,536 | 1.41 GiB | 0.19 GiB | 7.54× |
| 131,072 | 2.81 GiB | 0.31 GiB | 9.03× |