Model comparison
ArrowMint-Gemma3-4B-YUKI-v0.1 vs Llama-3.1-8B-Instruct
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: ArrowMint-Gemma3-4B-YUKI-v0.1'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
| ArrowMint-Gemma3-4B-YUKI-v0.1 | Llama-3.1-8B-Instruct | |
|---|---|---|
| Parameters | 4.3B | 8.0B |
| Architecture | gemma3 | llama |
| Layers | 34 | 32 |
| Native context | 131,072 | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 24 | 45 |
| Smallest quantization | 1.06 GiB | 2.02 GiB |
| Q4_K_M | — | 4.58 GiB |
| Licence | gemma | llama3.1 |
KV cache by context
the term that decides long-context viability
| Context | ArrowMint-Gemma3-4B-YUKI-v0.1 | Llama-3.1-8B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.25 GiB | 0.50 GiB | 2.02× |
| 8,192 | 0.33 GiB | 1.00 GiB | 3.07× |
| 16,384 | 0.48 GiB | 2.00 GiB | 4.15× |
| 32,768 | 0.79 GiB | 4.00 GiB | 5.03× |
| 65,536 | 1.42 GiB | 8.00 GiB | 5.63× |
| 131,072 | 2.67 GiB | 16.00 GiB | 5.99× |