Updates
3/10/2026
I've uploaded new quants using the new fused Up + Gate conversion, this offers up to a +10% boost in prompt processing speed from my testing.
Description
This repo contains specialized MoE-quants for Qwen3.5-35B-A3B. 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 |
|---|---|---|---|---|---|
| Q5_K_M | 24.45 GiB (6.06 BPW) | Q8_0 / Q5_K / Q5_K / Q6_K | 6.534165 ± 0.041552 | -0.0182% | 0.006174 ± 0.000108 |
| Q4_K_M | 20.62 GiB (5.11 BPW) | Q8_0 / Q4_K / Q4_K / Q5_K | 6.564172 ± 0.041821 | +0.4409% | 0.010121 ± 0.000122 |
| IQ4_XS | 16.40 GiB (4.06 BPW) | Q8_0 / IQ3_S / IQ3_S / IQ4_XS | 6.632992 ± 0.042286 | +1.4940% | 0.024130 ± 0.000250 |
| IQ3_S | 12.65 GiB (3.14 BPW) | Q8_0 / IQ2_S / IQ2_S / IQ3_S | 6.920534 ± 0.044619 | +5.8937% | 0.061926 ± 0.000402 |
- Downloads last month
- 9,460
Hardware compatibility
Log In to add your hardware
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support