Model comparison
KaLM-embedding-multilingual-mini-instruct-v2.5 vs bge-m3
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: KaLM-embedding-multilingual-mini-instruct-v2.5's KV cache at 32K is 8.0× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| KaLM-embedding-multilingual-mini-instruct-v2.5 | bge-m3 | |
|---|---|---|
| Parameters | 494M | 567M |
| Architecture | qwen2 | bert |
| Layers | 24 | 24 |
| Native context | 131,072 | 8,194 |
| Mixture of experts | no | no |
| Quantizations published | 1 | 1 |
| Smallest quantization | 0.49 GiB | 0.59 GiB |
| Q4_K_M | — | — |
| Licence | apache-2.0 | mit |
KV cache by context
the term that decides long-context viability
| Context | KaLM-embedding-multilingual-mini-instruct-v2.5 | bge-m3 | Ratio |
|---|---|---|---|
| 4,096 | 0.05 GiB | 0.38 GiB | 8.00× |
| 8,192 | 0.09 GiB | 0.75 GiB | 8.00× |
| 16,384 | 0.19 GiB | 1.50 GiB | 8.00× |
| 32,768 | 0.38 GiB | 3.00 GiB | 8.00× |
| 65,536 | 0.75 GiB | 6.00 GiB | 8.00× |
| 131,072 | 1.50 GiB | 12.00 GiB | 8.00× |