Model comparison
grug-9b vs gemma-4-31B-it
At Q4_K_M, grug-9b is the smaller download — 5,629,109,056 bytes against 18,323,733,440. At long context the gap widens: grug-9b's KV cache at 32K is 6.2× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| grug-9b | gemma-4-31B-it | |
|---|---|---|
| Parameters | 9.4B | 31.3B |
| Architecture | qwen35 | gemma4 |
| Layers | 32 | 60 |
| Native context | 262,144 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 28 | 55 |
| Smallest quantization | 3.51 GiB | 7.95 GiB |
| Q4_K_M | 5.24 GiB | 17.07 GiB |
| Licence | mit | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | grug-9b | gemma-4-31B-it | Ratio |
|---|---|---|---|
| 4,096 | 0.13 GiB | 1.80 GiB | 14.36× |
| 8,192 | 0.25 GiB | 2.42 GiB | 9.68× |
| 16,384 | 0.50 GiB | 3.67 GiB | 7.34× |
| 32,768 | 1.00 GiB | 6.17 GiB | 6.17× |
| 65,536 | 2.00 GiB | 11.17 GiB | 5.59× |
| 131,072 | 4.00 GiB | 21.17 GiB | 5.29× |