Model comparison
MathCoder2-CodeLlama-7B vs HyperCLOVAX-SEED-Text-Instruct-1.5B
At Q4_K_M, HyperCLOVAX-SEED-Text-Instruct-1.5B is the smaller download — 1,133,974,368 bytes against 4,081,096,640.
From the file· summed bytes, KV per layer
Side by side
| MathCoder2-CodeLlama-7B | HyperCLOVAX-SEED-Text-Instruct-1.5B | |
|---|---|---|
| Parameters | 6.7B | 1.6B |
| Architecture | llama | llama |
| Layers | 32 | — |
| Native context | 16,384 | — |
| Mixture of experts | no | no |
| Quantizations published | 20 | 1 |
| Smallest quantization | 2.20 GiB | 1.06 GiB |
| Q4_K_M | 3.80 GiB | 1.06 GiB |
| Licence | apache-2.0 | other |
KV cache by context
the term that decides long-context viability
| Context | MathCoder2-CodeLlama-7B | HyperCLOVAX-SEED-Text-Instruct-1.5B | Ratio |
|---|---|---|---|
| 4,096 | 2.00 GiB | — | — |
| 8,192 | 4.00 GiB | — | — |
| 16,384 | 8.00 GiB | — | — |
| 32,768 | 16.00 GiB | — | — |
| 65,536 | 32.00 GiB | — | — |
| 131,072 | 64.00 GiB | — | — |