Model comparison
Voxtral-Mini-3B-2507 vs cohere-transcribe-03-2026
At Q4_K_M, cohere-transcribe-03-2026 is the smaller download — 1,558,162,944 bytes against 2,473,002,304.
From the file· summed bytes, KV per layer
Side by side
| Voxtral-Mini-3B-2507 | cohere-transcribe-03-2026 | |
|---|---|---|
| Parameters | 4.7B | 2.1B |
| Architecture | voxtral | cohere_asr |
| Layers | 30 | — |
| Native context | 131,072 | — |
| Mixture of experts | no | no |
| Quantizations published | 31 | 11 |
| Smallest quantization | 1.45 GiB | 1.41 GiB |
| Q4_K_M | 2.30 GiB | 1.45 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Voxtral-Mini-3B-2507 | cohere-transcribe-03-2026 | Ratio |
|---|---|---|---|
| 4,096 | 0.47 GiB | — | — |
| 8,192 | 0.94 GiB | — | — |
| 16,384 | 1.88 GiB | — | — |
| 32,768 | 3.75 GiB | — | — |
| 65,536 | 7.50 GiB | — | — |
| 131,072 | 15.00 GiB | — | — |