| MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. |
|
|
| This repository contains unofficial Model Derivatives and is not endorsed by MiniMax or Comfy-Org. |
|
|
| The repository owner confirms separate written MiniMax authorization covering this publication. That authorization is not sublicensed here; downstream users remain responsible for the official license and location-specific authorization. |
|
|
| Source lineage: |
| - Base model: MiniMaxAI/MiniMax-H3 revision 6818f6c32d12b210915e44ad56a4228c2608f160. |
| - Pruned FL2VA lineage: Comfy-Org/MiniMax-H3 revision 014cd40f7e177756c6b2473c0d93b1c89a790dd2. |
| - Converter: unslothai/stable-diffusion.cpp revision 13b9d92b5e9a1563536c9c980e700470f9ab6702. |
|
|
| Original uniform derivatives, now deprecated for perceptual failure: |
| - minimax_h3_fl2va_pruned-IQ1_S.gguf and minimax_h3_fl2va_pruned-IQ1_M.gguf were converted from the pruned FL2VA BF16 denoiser to GGUF and quantized using an H3 activation importance matrix. |
| - Norm weights and condition_proj.weight were retained in BF16. |
| - These files passed structural loading and container smoke checks but failed the later controlled perceptual gate. They remain only for reproducibility and are not recommended for generation. |
|
|
| Final QF2 importance-guided mixed-precision derivatives: |
| - minimax_h3_fl2va_pruned-UD-IQ1_S-QF.gguf and minimax_h3_fl2va_pruned-UD-IQ1_M-QF.gguf were requantized from minimax_h3_fl2va_pruned-Q8_0.gguf in unsloth/MiniMax-H3-GGUF revision 9ee8213df85a2fcec53dec8c651a0fb1e821674a. |
| - The activation importance matrix was refreshed with eight representative Q8 audiovisual generation runs. It is 18,693,323 bytes and its SHA-256 is 6ce80c23f85d50cc170b1bfafa681811e5ef26590757f1c31540485b502c67a2. |
| - condition_proj.weight and all one-dimensional norm/gain tensors are retained at BF16/source precision. |
| - Both token-refiner blocks and every eligible matrix in main transformer blocks 46-49 use Q8_0. |
| - In main blocks 30-45, mlp.fc1 uses Q4_K. |
| - Attention qkv/out and mlp.fc2 use Q4_K in all main blocks except blocks 46-49, which are captured first by the Q8_0 rule. |
| - Only mlp.fc1 in blocks 0-29 uses IQ1_S or IQ1_M according to the output file. |
| - Audio/video patch-projection weights remain Q8_0 from the Q8 conversion input because their shapes are incompatible with IQ1/Q4_K conversion. condition_proj.weight is separately forced to BF16. |
| - Biases, final/output layers, and other incompatible tensor shapes remain at source precision under converter safeguards. |
| - These are mixed-precision derivatives whose lowest precision is IQ1; they are not uniform one-bit models. |
|
|
| Final QF2 artifacts: |
| - minimax_h3_fl2va_pruned-UD-IQ1_S-QF.gguf: 10,876,753,440 bytes; SHA-256 b12244f203e257f10c9cec56b8062aa65d627d9250c433ef633a244d610b292c. |
| - minimax_h3_fl2va_pruned-UD-IQ1_M-QF.gguf: 10,985,133,600 bytes; SHA-256 c7993a8a0eb202f63f30cb410817fc0a504f3c844ea4e8dbb38599d29b0d4bdc. |
| - Each contains 532 tensors: BF16 212, F16 102, F32 8, Q8_0 26, Q4_K 154, and IQ1 30. The IQ1 subtype follows the filename. |
| - Each contains 20,111,438,744 parameters: BF16 28,113,664; F16 43,642,368; F32 707,224; Q8_0 2,312,798,208; Q4_K 13,101,957,120; IQ1 4,624,220,160. |
|
|
| Validation scope and limitation: |
| - Both QF2 variants pass the four-step 640x384, 39-frame fox website profile with clear video and healthy audio. |
| - Both QF2 variants also pass an independent four-step motorcycle holdout. |
| - Both produce excellent visual results on the eight-step fox profile, but both have materially quieter audio than the matching Q8 control and therefore fail the eight-step audio gate. |
| - Motorcycle-prompt clipping is disclosed relative to the Q8 baseline in QUALITY_REPORT.md; no clipping-free claim is made. |
|
|
| Earlier internal QF1 working builds are superseded and are not final repository artifacts. Their identities and limited results are retained only in QUALITY_REPORT.md for audit history. |
|
|
| All conversions and quantizations change model numerics. Final file sizes, SHA-256 hashes, tensor inventories, and controlled validation results are recorded in README.md, QUALITY_REPORT.md, BUILD_MANIFEST.json, and SHA256SUMS. |
|
|
| LICENSE is the MiniMax H3 Community License Agreement supplied with the base model. Read it before use; the separate publication authorization described above is not transferred to downstream users. |
|
|
