--- base_model: - Qwen/Qwen3.6-35B-A3B --- ## Updates - 05-18-26: - ~~I have re-uploaded all quants with MTP tensors (@ Q8_0) now included.~~ - ~~Edit: I forgot to do `--fuse-gate-up-exp` so the quants are missing imatrix for some weights. Redoing and re-uploading. Will update when complete.~~ - Fixed, quants are now properly fused + imatrixed + have MTP! ## Description This repo contains specialized MoE-quants for Qwen3.6-35B-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 | | :----- | :------------------- | :---------------------------- | :------------------ | :------------------------ | :------------------ | | Q8_0 | 35.19 GiB (8.51 BPW) | Q8_0 | 6.722728 ± 0.043683 | +0.0412% | 0.005988 ± 0.000117 | | Q6_K | 27.93 GiB (6.76 BPW) | Q8_0 / Q6_K / Q6_K / Q6_K | 6.719363 ± 0.043660 | -0.0089% | 0.006522 ± 0.000094 | | Q5_K_M | 25.28 GiB (6.12 BPW) | Q8_0 / Q5_K / Q5_K / Q6_K | 6.731999 ± 0.043789 | +0.1791% | 0.008319 ± 0.000134 | | Q4_K_M | 21.45 GiB (5.19 BPW) | Q8_0 / Q4_K / Q4_K / Q5_K | 6.742029 ± 0.043830 | +0.3284% | 0.013652 ± 0.000131 | | IQ4_XS | 17.23 GiB (4.17 BPW) | Q8_0 / IQ3_S / IQ3_S / IQ4_XS | 6.878696 ± 0.044917 | +2.3622% | 0.032913 ± 0.000260 | | IQ3_S | 13.48 GiB (3.26 BPW) | Q6_K / IQ2_S / IQ2_S / IQ3_S | 7.168835 ± 0.047360 | +6.6797% | 0.084335 ± 0.000565 | ![kld_graph](kld_data/01_kld_vs_filesize.png "Chart showing Pareto KLD analysis of quants") ![ppl_graph](kld_data/02_ppl_vs_filesize.png "Chart showing Pareto PPL analysis of quants")