| ▲ | embedding-shape 4 hours ago | |
Sounds like maybe using worse quantization on the bigger model? Quantization matters a lot for the quality, basically anything below Q8 is borderline unusable. If it isn't specified in a benchmark already it probably should. | ||