Model comparison
gpt2 vs TinyMistral-248M-v2-Instruct
At Q4_K_M, gpt2 is the smaller download — 112,858,624 bytes against 155,668,768.
From the file· summed bytes, KV per layer
Side by side
| gpt2 | TinyMistral-248M-v2-Instruct | |
|---|---|---|
| Parameters | 137M | 248M |
| Architecture | gpt2 | llama |
| Layers | 12 | 12 |
| Native context | — | 32,768 |
| Mixture of experts | no | no |
| Quantizations published | 49 | 6 |
| Smallest quantization | 0.06 GiB | 0.10 GiB |
| Q4_K_M | 0.11 GiB | 0.14 GiB |
| Licence | — | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | gpt2 | TinyMistral-248M-v2-Instruct | Ratio |
|---|---|---|---|
| 4,096 | — | — | — |
| 8,192 | — | — | — |
| 16,384 | — | — | — |
| 32,768 | — | — | — |
| 65,536 | — | — | — |
| 131,072 | — | — | — |