This repo contains specialized MoE-quants for Step-3.5-Flash. The idea being that given the huge size of the FFN tensors compared to the rest of the tensors in the model, it should be possible to achieve a better quality while keeping the overall size of the entire model smaller compared to a similar naive quantization. To that end, the quantization type default is kept in high quality and the FFN UP + FFN GATE tensors are quanted down along with the FFN DOWN tensors.
| Quant | Size | Mixture | PPL | 1-(Mean PPL(Q)/PPL(base)) | KLD |
|---|---|---|---|---|---|
| Q4_K_M | 113.82 GiB (4.96 BPW) | Q8_0 / Q4_K / Q4_K / Q5_K | 4.718049 ± 0.030373 | +0.3762% | 0.015464 ± 0.000133 |
| IQ4_XS | 88.90 GiB (3.88 BPW) | Q8_0 / IQ3_S / IQ3_S / IQ4_XS | 4.822499 ± 0.031236 | +2.5984% | 0.042753 ± 0.000301 |
| IQ3_XXS | 73.10 GiB (3.19 BPW) | Q6_K / IQ3_XXS / IQ3_XXS / IQ3_XXS | 4.882908 ± 0.031560 | +3.8836% | 0.078681 ± 0.000506 |
- Downloads last month
- 2,060
Hardware compatibility
Log In to add your hardware
3-bit
4-bit
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for AesSedai/Step-3.5-Flash-GGUF
Base model
stepfun-ai/Step-3.5-Flash