---
base_model:
- MiniMaxAI/MiniMax-M3
---
This repo contains specialized MoE-quants for MiniMaxAI/MiniMax-M3. 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.

Additionally, I've provided a [J-Space lens](https://transformer-circuits.pub/2026/workspace/index.html) in the `lens/` folder, trained on ~128 text prompts.

| Quant  | Size                  | Mixture                          | PPL                 | 1-(Mean PPL(Q)/PPL(base)) | KLD                 |
| :----- | :-------------------- | :------------------------------- | :------------------ | :------------------------ | :------------------ |
| Q8_0   | 421.84 GiB (8.50 BPW) | Q8_0                             | 5.202737 ± 0.034827 | +0.3357%                  | 0.030136 ± 0.000884 |
| Q5_K_M | 295.20 GiB (5.95 BPW) | Q8_0 / Q5_K / Q5_K / Q6_K        | 5.218907 ± 0.034969 | +0.6475%                  | 0.041986 ± 0.001016 |
| Q4_K_M | 246.11 GiB (4.96 BPW) | Q8_0 / Q4_K / Q4_K / Q5_K        | 5.282848 ± 0.035408 | +1.8807%                  | 0.068863 ± 0.000982 |
| IQ4_XS | 189.12 GiB (3.81 BPW) | Q6_K / IQ3_S / IQ3_S / IQ4_XS    | 5.657391 ± 0.038606 | +9.1038%                  | 0.151112 ± 0.001437 |
| IQ3_S  | 148.04 GiB (2.98 BPW) | Q6_K / IQ2_S / IQ2_S / IQ3_S     | 6.417468 ± 0.044920 | +23.7620%                 | 0.321256 ± 0.002052 |
| IQ2_S  | 134.01 GiB (2.70 BPW) | Q6_K / IQ2_XS / IQ2_XS / IQ3_XXS | 6.636852 ± 0.046125 | +27.9929%                 | 0.397892 ± 0.002336 |

![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")