Model comparison
starcoder2-3b vs Llama-3.1-8B-Instruct
At Q4_K_M, starcoder2-3b is the smaller download — 1,848,976,448 bytes against 4,920,739,168.
From the file· summed bytes, KV per layer
Side by side
| starcoder2-3b | Llama-3.1-8B-Instruct | |
|---|---|---|
| Parameters | 3.0B | 8.0B |
| Architecture | starcoder2 | llama |
| Layers | 30 | 32 |
| Native context | 16,384 | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 37 | 45 |
| Smallest quantization | 1.07 GiB | 2.02 GiB |
| Q4_K_M | 1.72 GiB | 4.58 GiB |
| Licence | bigcode-openrail-m | llama3.1 |
KV cache by context
the term that decides long-context viability
| Context | starcoder2-3b | Llama-3.1-8B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | — | 0.50 GiB | — |
| 8,192 | — | 1.00 GiB | — |
| 16,384 | — | 2.00 GiB | — |
| 32,768 | — | 4.00 GiB | — |
| 65,536 | — | 8.00 GiB | — |
| 131,072 | — | 16.00 GiB | — |