Model comparison
Supra-Router-51M vs gpt2
At Q4_K_M, Supra-Router-51M is the smaller download — 37,446,208 bytes against 112,858,624.
From the file· summed bytes, KV per layer
Side by side
| Supra-Router-51M | gpt2 | |
|---|---|---|
| Parameters | 52M | 137M |
| Architecture | llama | gpt2 |
| Layers | 12 | 12 |
| Native context | 5,120 | — |
| Mixture of experts | no | no |
| Quantizations published | 18 | 49 |
| Smallest quantization | 0.02 GiB | 0.06 GiB |
| Q4_K_M | 0.03 GiB | 0.11 GiB |
| Licence | apache-2.0 | — |
KV cache by context
the term that decides long-context viability
| Context | Supra-Router-51M | gpt2 | Ratio |
|---|---|---|---|
| 4,096 | 0.05 GiB | — | — |
| 8,192 | 0.09 GiB | — | — |
| 16,384 | 0.19 GiB | — | — |
| 32,768 | 0.38 GiB | — | — |
| 65,536 | 0.75 GiB | — | — |
| 131,072 | 1.50 GiB | — | — |