NOTICE
| 1 | unsloth/MiniMax-H3-GGUF |
| 2 | |
| 3 | "MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, |
| 4 | Copyright (c) 2026 MiniMax. All Rights Reserved." |
| 5 | |
| 6 | The weights here derive from MiniMax-H3 (https://huggingface.co/MiniMaxAI/MiniMax-H3). |
| 7 | LICENSE is the text MiniMax ships at |
| 8 | https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE, verbatim. |
| 9 | |
| 10 | Modified files and the changes made to them (MiniMax H3 Community License Agreement, |
| 11 | Section III.2): |
| 12 | |
| 13 | * minimax_h3_fl2va_pruned-*.gguf: the H3-Base FL2VA diffusion transformer, converted |
| 14 | from safetensors to GGUF and quantized at six uniform precisions (Q2_K, Q3_K, Q4_K, |
| 15 | Q5_0, Q6_K, Q8_0) and two mixed-precision ones (UD-Q2_K_XL, UD-Q3_K_XL), which |
| 16 | changes the numerics. The modulation path is the pruned form published by Comfy-Org |
| 17 | (https://huggingface.co/Comfy-Org/MiniMax-H3) rather than the dense original. The |
| 18 | 1-D norm scales and condition_proj are held at bf16 in every build. |
| 19 | * qwen3vl_32b_minimax_h3-*.gguf: the Qwen3-VL-32B text encoder MiniMax ship with H3, |
| 20 | converted to GGUF and quantized at Q2_K_M and Q4_K_M, which changes the numerics. |
| 21 | That encoder is Qwen3-VL-32B (https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct), |
| 22 | Apache-2.0, as redistributed by MiniMax inside the H3 repository. |
| 23 | |
| 24 | Not an official MiniMax product, and not endorsed by MiniMax. |
| 25 | |