metadata
language: en
tags:
- quantized
- mlx
base_model:
- moonshotai/Kimi-K2.5
library_name: mlx
pipeline_tag: text-generation
See Kimi-K2.5 MLX in action - demonstration video
Tested on a M3 Ultra 512GB RAM using Inferencer app v1.9.4
- Single inference ~26.82 tokens/s @ 1000 tokens
- Batched inference ~39 total tokens/s across three inferences
- Memory usage: ~440 GiB
q3.6bit quant typically achieves 1.59375 perplexity in our coding test
| Quantization | Perplexity | Token Accuracy | Missed Divergence |
|---|---|---|---|
| q3.5 | 168.0 | 43.45% | 72.57% |
| q3.6 | 1.59375 | 83.75% | 41.11% |
| q3.8 | 1.49218 | 86.65% | 36.79% |
| q4.2 | 1.39062 | 89.30% | 31.70% |
| q4.5 | 1.33593 | 91.65% | 27.61% |
| q6.5 | 1.21875 | 96.95% | 12.03% |
| Base | 1.20312 | 100.0% | 0.000% |
- Perplexity: Measures the confidence for predicting base tokens (lower is better)
- Token Accuracy: The percentage of correctly generated base tokens
- Missed Divergence: Measures severity of misses; how much the token was missed by
Quantized with a modified version of MLX
For more details see demonstration video or visit Kimi-K2.5.
Disclaimer
We are not the creator, originator, or owner of any model listed. Each model is created and provided by third parties. Models may not always be accurate or contextually appropriate. You are responsible for verifying the information before making important decisions. We are not liable for any damages, losses, or issues arising from its use, including data loss or inaccuracies in AI-generated content.