Model comparison
Tini-Cybersec-8B-A1B vs llama-3-youko-8b
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Tini-Cybersec-8B-A1B's KV cache at 32K is 10.7× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Tini-Cybersec-8B-A1B | llama-3-youko-8b | |
|---|---|---|
| Parameters | 8.5B | 8.0B |
| Architecture | lfm2moe | llama |
| Layers | 24 | 32 |
| Native context | 128,000 | 8,192 |
| Mixture of experts | yes, 32 experts | no |
| Quantizations published | 35 | 2 |
| Smallest quantization | 2.71 GiB | 5.34 GiB |
| Q4_K_M | 4.80 GiB | — |
| Licence | apache-2.0 | llama3 |
KV cache by context
the term that decides long-context viability
| Context | Tini-Cybersec-8B-A1B | llama-3-youko-8b | Ratio |
|---|---|---|---|
| 4,096 | 0.05 GiB | 0.50 GiB | 10.66× |
| 8,192 | 0.09 GiB | 1.00 GiB | 10.66× |
| 16,384 | 0.19 GiB | 2.00 GiB | 10.66× |
| 32,768 | 0.38 GiB | 4.00 GiB | 10.67× |
| 65,536 | 0.75 GiB | 8.00 GiB | 10.67× |
| 131,072 | 1.50 GiB | 16.00 GiB | 10.67× |