Model comparison
tada-3b-ml vs s2-pro
These two publish different quantization sets; the table below has the exact sizes.
From the file· summed bytes, KV per layer
Side by side
| tada-3b-ml | s2-pro | |
|---|---|---|
| Parameters | 4.2B | 4.6B |
| Architecture | wav2vec2 | fish-speech |
| Layers | 28 | 36 |
| Native context | 131,072 | — |
| Mixture of experts | no | yes, 1 experts |
| Quantizations published | 3 | 7 |
| Smallest quantization | 5.20 GiB | 2.40 GiB |
| Q4_K_M | — | 3.32 GiB |
| Licence | llama3.2 | other |
KV cache by context
the term that decides long-context viability
| Context | tada-3b-ml | s2-pro | Ratio |
|---|---|---|---|
| 4,096 | 0.44 GiB | — | — |
| 8,192 | 0.88 GiB | — | — |
| 16,384 | 1.75 GiB | — | — |
| 32,768 | 3.50 GiB | — | — |
| 65,536 | 7.00 GiB | — | — |
| 131,072 | 14.00 GiB | — | — |