Add MiniMax-H3 NVFP4 quantization study card, figures and comparison videos

Model card with measured recipe sweep, checkpoint comparison and local RTX 5090
benchmarks; MiniMax H3 Community License copy and NOTICE.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

LICENSE.MiniMax-H3

@@ -0,0 +1,84 @@
+MiniMax H3 COMMUNITY LICENSE AGREEMENT
+MiniMax H3 release date/License date: August 2, 2026.
+The scope of this License Agreement (this “Agreement”) is expressly limited to the “Applicable Territory” as defined below.
+By clicking to accept, or by using, reproducing, modifying, distributing, running, or displaying any portion or element of the MiniMax H3 Works (including through any Hosted Services) in any manner, you acknowledge and accept the terms of this Agreement, and this Agreement shall take immediate effect upon the occurrence of such act.
+I.  Definitions
+1. “Acceptable Use Policy” means the policy published by MiniMax in Exhibit A.
+2. “Agreement” means the terms and conditions set forth herein that govern the use, reproduction, distribution, modification, running, and display of the MiniMax H3 Works or any portion or element thereof.
+3. “Applicable Territory” means worldwide, excluding the Excluded Territories.
+4. “Documentation” means the specifications, manuals, and documentation concerning MiniMax H3 that are publicly released by MiniMax.
+5. “Excluded Territories” means the European Union, the United Kingdom, the Republic of Korea and the United States of America.  
+6. “MiniMax H3” means the video generation model, together with its software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-supporting code, and other elements thereof made publicly available by Us, as released at https://huggingface.co/MiniMaxAI/MiniMax-H3.
+7. “MiniMax H3 Works” means (i) the Materials, (ii) the Model Derivatives, and (iii) all derivatives thereof.
+8. “Hosted Services” means hosted services provided via application programming interfaces (APIs), web access, or any other electronic or remote means.
+9. “Licensee,” “you,” or “your” means the natural or legal person exercising rights and/or using the MiniMax H3 Works for any purpose in any field of use under this Agreement.
+10. “Materials” means, collectively, MiniMax H3 and the Documentation (and any portion thereof), in each case as made available by MiniMax under this Agreement and proprietary to MiniMax.
+11. “Model Derivatives” means all of the following: (i) any modification of MiniMax H3 or any Model Derivative thereof; (ii) any work based on MiniMax H3 or any Model Derivative thereof; or (iii) any other machine learning model created by transferring the patterns of the weights, parameters, operational patterns, or Outputs of MiniMax H3 or any Model Derivative thereof to another model, such that the latter model exhibits behavior similar to MiniMax H3 or its Model Derivatives, including by distillation methods, methods using intermediate data representations, or methods based on training using synthetic-data Outputs generated by MiniMax H3 or its Model Derivatives. For the avoidance of doubt, Outputs are not deemed Model Derivatives.
+12. “Output” means any result of operating or otherwise using MiniMax H3 or any Model Derivatives (including through Hosted Services).
+13. “Third Party” means any natural or legal person that is not under common control with us or with you.
+14. “Including” means “including but not limited to.”
+15. “We,” “Us”  or  “MiniMax” means Nanonoble Pte. Ltd..
+II.  Grant of Rights
+Solely within the Applicable Territory, we grant you a non-exclusive, non-transferable, royalty-free, limited license to use, reproduce, distribute, create derivative works (including Model Derivatives), and modify the Materials in accordance with the terms of this Agreement and the Acceptable Use Policy, based on the intellectual property and other rights owned by MiniMax that are embodied in or used by the Materials. You shall not violate (or encourage or permit any person to violate) any term of this Agreement or the Acceptable Use Policy.
+We will continuously evaluate the applicable laws, regulations and compliance requirements for the Excluded Territories. In the meantime, should any person in such Excluded Territories be interested in deploying our models, you are welcome to contact us about obtaining a license, which will be granted based on robust controls and guardrails for purposes of complying with the laws, regulations and compliance requirements of the Excluded Territories.
+III.  Distribution and Redistribution
+Subject to and conditioned on your continuing compliance with this Agreement, including its territorial restrictions and the Acceptable Use Policy, and solely within the Applicable Territory, you may distribute or make available the MiniMax H3 Works to Third Parties within the Applicable Territory; provided, that all of the following conditions are met:
+1. You must provide a copy of this Agreement to all such Third Parties who receive the MiniMax H3 Works or use your products or services related thereto;
+2. You must cause any modified files to carry prominent notices stating that you have modified such files;
+3. You are encouraged to:
+  a. display a notice on any product or service developed using MiniMax H3 indicating that the product or service is “Powered by MiniMax H3”; 
+  b. add an AI-generation identifier to files produced using generative AI models including MiniMax H3; and
+  c. publish at least one technical blog post or a public statement describing your experience using MiniMax H3 Works;
+4. All distributions to Third Parties (other than through Hosted Services) must be accompanied by a “NOTICE” text file containing the following notice:
+  “MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.”
+You may add your own copyright notices on your modifications; except as provided in this Section and in Section V, however, you may not impose additional or different terms and conditions on the use, reproduction, or distribution of your modifications or of any aggregate Model Derivatives, and your use, reproduction, modification, distribution, running, and display of the work must otherwise comply with the terms and conditions of this Agreement (including the provisions concerning the Applicable Territory). If you receive the MiniMax H3 Works from a Licensee as part of an integrated end-user product, the provisions of Section III of this Agreement do not apply to you, but Section V and Exhibit A remain applicable.
+IV.  Additional Commercial Terms
+1. You shall obtain a separate, prior written authorization from MiniMax by contacting api@minimax.io with the subject line “MiniMax H3 licensing - authorization request”, if your commercial products and services generate more than 20 million US dollars (or equivalent in other currencies) in yearly revenue.
+2. You shall prominently display “MiniMax H3”on the user interface of commercial product or service that uses MiniMax H3 or MiniMax H3 Works.
+V. Use Restrictions
+1. Your use of the MiniMax H3 Works must comply with applicable laws and regulations (including trade-compliance laws and regulations) and must comply with the Acceptable Use Policy for the MiniMax H3 Works, which is incorporated into this Agreement by reference. 
+2. Before providing access to the MiniMax H3 Works or any product, service, or Hosted Service incorporating them, you must bind each recipient or user to enforceable terms at least as protective as the use restrictions in this Section V and Exhibit A, and you must notify each recipient or user that those restrictions apply.
+3. You may not use the MiniMax H3 Works or any of their Outputs or results to improve any other artificial intelligence model (other than MiniMax H3 or its Model Derivatives).
+4. You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory. Any such use outside the Applicable Territory is not authorized by this Agreement.
+5. If you provide or make available to any Third Party a product, service, or Hosted Service that permits the generation of Outputs using MiniMax H3 or any Model Derivative, you must, before making that product or service available and throughout its operation, implement, maintain, test, and periodically review reasonable and proportionate technical and organizational safeguards designed to prevent and mitigate access, uses, and Outputs that violate this Section V or Exhibit A, including uses or Outputs that infringe, misappropriate, or otherwise violate any Third Party’s intellectual-property or other rights. You must not knowingly disable, materially weaken, or permit the circumvention of those safeguards. You must maintain a reasonably accessible mechanism for reporting suspected violations. Upon receiving a good-faith report or otherwise obtaining actual knowledge of a violation, you must promptly investigate and take reasonable steps within your control to stop or mitigate the violation, including removing or disabling access to offending content or services and suspending or terminating repeat violators where appropriate. You are responsible for implementing and enforcing these requirements with respect to your products, services, systems, users, and downstream recipients.
+VI.  Intellectual Property
+1. Subject to MiniMax’s rights in the MiniMax H3 Works (and the intellectual property therein), and to your compliance with the terms and conditions of this Agreement, as between you and MiniMax, you will own the derivative works and modifications of the Materials that you have created or had created, as well as any Model Derivatives.
+2. Except for the limited license expressly granted in this paragraph, no trademark license is granted under this Agreement; with respect to MiniMax H3 Works, the Licensee may not use any name or mark owned by or associated with MiniMax or any of its affiliates, except as reasonably and customarily necessary to describe and distribute the MiniMax H3 Works. MiniMax hereby grants you a license to use the “MiniMax H3” mark (the “Mark”) within the Applicable Territory solely for the purpose of complying with Section III.3; provided, that you comply with all applicable trademark-protection laws. All goodwill arising from your use of the Mark shall inure to the benefit of MiniMax.
+3. If you bring or assert any suit or other legal proceeding (including a cross-claim or counterclaim in any action) against us or any other natural or legal person alleging that the Materials, any Output, or any portion of the foregoing infringes any intellectual property right or other right owned by you or for which you can obtain a license, all licenses granted to you under this Agreement will terminate as of the date such suit or proceeding is filed. You shall defend, indemnify, and hold us harmless against any Third-Party claim arising out of or related to the use or distribution of the MiniMax H3 Works by you or by any Third Party.
+4. MiniMax claims no rights over the Outputs you generate. You and your users are entirely responsible for the Outputs and any subsequent use thereof.
+VII.  Disclaimers and Limitations of Liability
+1. We have no obligation to support, update, provide training for, or develop any further version of the MiniMax H3 Works, or to grant any license with respect thereto.
+2. UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE MINIMAX H3 WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE MINIMAX H3 WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE MINIMAX H3 WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT.
+3. TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL MINIMAX OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE MINIMAX H3 WORKS OR OUTPUTS, EVEN IF MINIMAX OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
+VIII.  Term and Termination
+1. This Agreement is effective from the moment you accept this Agreement or begin accessing the Materials, and, subject to your compliance with its terms and conditions, will remain in effect until terminated as provided herein.
+2. If you breach any term or condition of this Agreement, we have the right to terminate this Agreement. Upon termination, you must immediately cease accessing, using, and distributing the MiniMax H3 Works; delete or destroy all copies within your possession or control; and notify each downstream recipient that your authorization has ended. The obligations in the preceding sentence and Sections VI.1, VI.3, VII, and IX survive termination.
+IX.  Governing Law and Jurisdiction
+1. This Agreement, and any dispute arising out of or related to this Agreement, shall be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to its conflict-of-laws rules. The United Nations Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
+2. Any dispute arising out of or related to this Agreement shall be subject to the exclusive jurisdiction of the courts of the Hong Kong Special Administrative Region of the People’s Republic of China with competent jurisdiction. Both MiniMax and the Licensee hereby consent to the exclusive jurisdiction of such courts for any such dispute.
+Additional Note: Please note that the encoder of MiniMax H3 uses Qwen3-VL-32B, which is licensed under Apache 2.0 License: https://github.com/QwenLM/Qwen3-VL/blob/main/LICENSE.
+
+Exhibit A — Acceptable Use Policy
+MiniMax reserves the right to update this Acceptable Use Policy from time to time. 
+Last revised: August 2, 2026.
+MiniMax is committed to promoting the safe and fair use of its tools and features, including MiniMax H3. You agree not to use MiniMax H3, any Model Derivatives, or any Output in any of the following ways:
+1. Use outside the Applicable Territory;
+2. Use in any manner that violates any applicable national, federal, state, local, or international law, regulation, or other legal requirement, or that infringes, misappropriates, or otherwise violates any Third Party’s intellectual-property or other proprietary rights, including through unauthorized reproduction, distribution, public display, public performance, or creation of derivative works;
+3. Use in any manner that may harm yourself or others;
+4. Use to repurpose or distribute the Outputs of MiniMax H3 or any Model Derivatives in order to harm yourself or others;
+5. Use to circumvent or bypass any safety guardrails or safeguards we have implemented;
+6. Use in any manner that exploits or harms, or intends to exploit or harm, minors;
+7. Use to generate or disseminate verifiably false information and/or content for the purpose of harming others or influencing elections;
+8. Use to manufacture or facilitate false online engagement, including fake reviews and other means of false online engagement;
+9. Use to intentionally defame, disparage, or otherwise harass others;
+10. Use to generate and/or disseminate malware (including ransomware) or any other content intended to damage electronic systems;
+11. Use to generate or disseminate personally identifiable information for the purpose of harming others;
+12. Use to generate or disseminate information (including images, code, posts, or articles) in or to any public environment (including via bot tweets or similar means) without clearly and prominently disclosing that such information and/or content is machine-generated;
+13. Use to impersonate another person without that person’s consent, authorization, or lawful right to do so;
+14. Use to make high-risk automated decisions in critical domains that affect individual safety, rights, or well-being (such as law enforcement, immigration, healthcare or medical services, critical-infrastructure management, product-safety components, essential services, credit, employment, housing, education, social scoring, or insurance);
+15. Use in any manner that violates or disregards the social, ethical, or moral standards of other countries or regions;
+16. Use to carry out, assist, threaten, incite, plan, advocate for, or encourage violent extremism or terrorism;
+17. Use for any purpose intended to discriminate against, or harm, individuals or groups based on protected characteristics or categories, online or offline social behavior, or known or predicted personality traits;
+18. Use to intentionally exploit the vulnerabilities of specific populations based on age, social, physical, or psychological characteristics, so as to materially distort the behavior of a member of that group in a manner that causes, or is likely to cause, physical or psychological harm to that person or to others;
+19. Use for military purposes;
+20. Use to engage in any unauthorized or unlicensed professional activity, including but not limited to financial, legal, medical or healthcare, or other professional practice.

NOTICE

@@ -0,0 +1,10 @@
+MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.
+
+The files in this repository are modified (quantized) derivatives of MiniMax-H3
+(https://huggingface.co/MiniMaxAI/MiniMax-H3):
+  - The transformer's per-block AdaLN projections are replaced by an equivalent rank-12 factorization
+    (ComfyUI curve form: `adaln_t_table` plus per-block projections).
+  - Transformer GEMM weights are quantized to NVFP4 and/or FP8 with NVIDIA Model Optimizer, and keys are
+    renamed and fused into ComfyUI's layout.
+
+The MiniMax-H3 text encoder is based on Qwen3-VL-32B, licensed under the Apache License 2.0.

README.md

@@ -0,0 +1,212 @@
+---
+license: other
+license_name: minimax-h3-community-license
+license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
+base_model: MiniMaxAI/MiniMax-H3
+base_model_relation: quantized
+pipeline_tag: text-to-video
+tags:
+  - nvfp4
+  - fp8
+  - quantization
+  - modelopt
+  - comfyui
+  - video-generation
+  - audio-generation
+---
+
+# MiniMax-H3: NVFP4 / FP8 quantization study
+
+> **Status: work in progress.** The measurements below are final for what they cover; the checkpoints and the
+> local ComfyUI comparisons (speed, VRAM, side-by-side clips on an RTX 5090) are still being produced.
+
+A measured look at quantizing the [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) video+audio transformer
+with [NVIDIA Model Optimizer](https://github.com/NVIDIA/Model-Optimizer) (ModelOpt), and checkpoints that load in
+stock ComfyUI. Every number on this page comes from a results file and is regenerated by one script.
+
+## TL;DR
+
+- **13 of H3's 33.1B transformer parameters are AdaLN projections that depend only on the timestep.** They factor
+  exactly into a rank-12 basis (66 GB → ~40 GB in BF16) at any step count and shift, including audio shift 2.
+  ComfyUI's pruned build uses the same idea at rank 8; at the output the two are indistinguishable.
+- **4-bit weights, not 4-bit activations, carry most of the NVFP4 error.** Weight-only NVFP4 is about as far from
+  BF16 as full W4A4, so methods that fix weight rounding matter most.
+- **Calibration helps, modestly.** Local-Hessian scales cut the per-step error of a naive NVFP4 cast by ~14%.
+- **Precision is best spent on attention.** FP8 attention projections with NVFP4 feed-forward roughly double the
+  quality bought per extra GB compared with keeping the first and last blocks in BF16.
+- **Block-wise scale learning did not help here:** it overfit the calibration prompts and made held-out error worse.
+- **Against published NVFP4 files:** at the same 12.5–12.7 GB, our calibrated file has ~15% less video error than an
+  uncalibrated cast; a 20 GB file that keeps attention in FP8 is better still, especially for audio.
+- **On an RTX 5090 attention dominates each step (~73%)**, so NVFP4 alone is only ~9% faster than INT8; a per-layer
+  attention config LoRA (kitchen INT8 / SOL sparse attention) gives ~1.8×.
+
+## What's in the transformer
+
+![Parameter breakdown](assets/params.png)
+
+<!-- BEGIN:params -->
+| Part | Parameters |
+|---|---|
+| AdaLN projections | 13.01B |
+| Attention | 7.71B |
+| Feed-forward | 11.56B |
+| Other | 0.84B |
+| **Total** | **33.12B** |
+<!-- END:params -->
+
+The AdaLN projections map `silu(temb(t))` — a function of the scalar timestep only — to every block's
+shift/scale/gate table. The timestep enters a 256-channel sinusoid unscaled in [0, 1], so the curve it traces
+is extremely smooth: its singular values fall to 1e-7 of the first by rank 12.
+
+![AdaLN error vs rank](assets/adaln_rank.png)
+
+<!-- BEGIN:adaln -->
+| Rank | Max rel. error | Mean rel. error |
+|---|---|---|
+| 4 | 3.1e-01 | 1.2e-03 |
+| 6 | 6.0e-02 | 1.5e-04 |
+| 8 | 4.4e-03 | 1.3e-05 |
+| 10 | 1.2e-03 | 3.5e-06 |
+| 12 | 1.2e-04 | 4.4e-07 |
+| 16 | 1.2e-04 | 4.3e-07 |
+| 24 | 1.2e-04 | 4.3e-07 |
+| 32 | 1.1e-04 | 4.4e-07 |
+| shipped bf16 path vs fp32 | 1.7e-01 | 1.5e-03 |
+<!-- END:adaln -->
+
+Both ComfyUI's rank-8 curve form and this rank-12 basis are orders of magnitude below the rounding error the
+shipped bf16 checkpoint already has. The rank-12 basis is kept for margin and because it allows a LoRA's AdaLN
+deltas to be folded in exactly (ComfyUI's pruned files store no basis, so such LoRA tensors are skipped).
+
+## Quantization recipes
+
+Each recipe quantizes only the 50 blocks' GEMMs (q/k/v fused, as ComfyUI stores them); AdaLN, the text refiner,
+embedders and output heads keep their shipped precision. Scores are the mean relative error of the predicted
+video/audio velocity against BF16, one denoising step at a time, on 8 held-out prompts (960×544, 124 frames).
+
+![Recipe comparison](assets/recipes.png)
+
+<!-- BEGIN:recipes -->
+| Recipe | Video error | Audio error | ≈ Size |
+|---|---|---|---|
+| Unquantized, factored AdaLN (noise floor) | 1.0% | 0.6% | 40 GB |
+| FP8 W8A8 (all GEMMs) | 7.3% | 6.4% | 21 GB |
+| NVFP4 FFN + FP8 attention | 13.3% | 10.8% | 16 GB |
+| NVFP4, blocks 0-1/48-49 BF16 | 15.5% | 14.1% | 15 GB |
+| NVFP4, local-Hessian scales | 15.8% | 14.9% | 13 GB |
+| NVFP4 weights only (W4A16, RTN) | 16.0% | 13.6% | 13 GB |
+| NVFP4, MSE scales | 16.9% | 15.6% | 13 GB |
+| NVFP4, max scales (naive cast) | 18.4% | 16.7% | 13 GB |
+| NVFP4, learned scales (overfit) \* | 21.2% | 16.2% | 13 GB |
+
+\* Separate run; its own unquantized floor was 1.7% video and its local-Hessian start point 15.6% video, so compare it to that start point.
+<!-- END:recipes -->
+
+![Quality vs size](assets/tradeoff.png)
+
+![Error across denoising steps](assets/steps.png)
+
+<!-- BEGIN:steps -->
+| Recipe | step 0 | step 4 | step 7 | step 11 | step 14 | step 18 |
+|---|---|---|---|---|---|---|
+| FP8 W8A8 (all GEMMs) | 6.1% | 13.6% | 10.5% | 6.0% | 4.0% | 3.5% |
+| NVFP4 FFN + FP8 attention | 12.3% | 23.3% | 17.8% | 10.1% | 7.6% | 8.5% |
+| NVFP4, local-Hessian scales | 15.4% | 28.4% | 20.6% | 11.9% | 8.8% | 9.6% |
+| NVFP4, max scales (naive cast) | 20.4% | 32.3% | 23.3% | 13.9% | 10.0% | 10.7% |
+<!-- END:steps -->
+
+### Block-wise scale learning: a negative result
+
+ModelOpt's `lsq` makes each NVFP4 block scale trainable (weights frozen). Tuning the 50 blocks one at a time from
+the local-Hessian start point, each block fit to the BF16 stream's output, lowered the error on the calibration
+prompts by 5–15% in the deep blocks, but **raised the held-out one-step error from 15.6% to 21.2% video** (audio
+15.0% → 16.2%). The quantized stream drifts far from BF16 by the deep blocks (~0.25 relative MSE at block 40), so
+asking each block's scales to undo that drift overfit the calibration prompts. A retry fitting each block only to
+its own error, with keep/revert decided on held-out prompts, ran out of memory and was not repeated; even before
+overfitting, the gains on early and middle blocks were 0.5–5%, so the ceiling looked low for this model.
+
+![Block-wise scale learning](assets/lsq_blocks.png)
+
+### Why one-step error and not video PSNR?
+
+H3's sampler is chaotic: computing only the AdaLN in fp32 instead of bf16 already moves a 20-step clip to
+~17 dB PSNR against the shipped model. End-of-run pixel metrics therefore cannot rank recipes. Each recipe is
+instead fed the exact recorded inputs of every step and compared with the BF16 output of that step.
+
+### Measurement notes
+
+- **Noise floor ~1–2%.** Replays are bitwise exact within one container but differ by 1–2% across GPU hosts, and
+  any bf16-level change lands at that level. Compare recipes within a run; each run reports its own floor.
+- **Size** is the approximate transformer checkpoint size with the rank-12 AdaLN (NVFP4 counted as 4.5 bits per
+  weight including block scales).
+- **Simulation vs ComfyUI.** NVFP4 activation rounding is implementation-defined: ComfyUI's CUDA and Triton kernels
+  agree with each other but differ from ModelOpt's simulation by ~2% on a quantized activation. Rankings should
+  carry over; final quality is judged in ComfyUI itself (below).
+
+## Published NVFP4 checkpoints, compared
+
+Each file's stored 4-bit weights are decoded exactly into the same model, activations are quantized the way ComfyUI
+does it (stored `input_scale`, else computed per call), and all are scored like the recipes above on the same
+held-out prompts in one run.
+
+<!-- BEGIN:ckpts -->
+| Checkpoint | Size | Video error | Audio error |
+|---|---|---|---|
+| BF16, factored AdaLN (noise floor) | 40.5 GB | 0.6% | 0.6% |
+| **This work**, NVFP4 local-Hessian | 12.7 GB | 15.4% | 15.3% |
+| lilcheaty `pruned_nvfp4` (same layout) | 12.5 GB | 18.2% | 16.8% |
+| rockerBOO `pruned_nvfp4` (FP8 qkv, BF16 out_proj + 5 edge blocks) | 20.1 GB | 12.7% | 9.1% |
+<!-- END:ckpts -->
+
+At the same size and speed, calibration lowers the error of an uncalibrated NVFP4 cast by ~15% (video). Keeping
+attention in FP8 and edge blocks in BF16 buys more quality, especially audio, for ~7 GB and ~15% speed.
+
+## Local results (ComfyUI, RTX 5090)
+
+<!-- BEGIN:local -->
+Stock ComfyUI (custom nodes disabled), NVIDIA GeForce RTX 5090, 960×544, 124 frames, PyTorch attention unless a config LoRA says otherwise. Median of 2 timed generations after a warm-up.
+
+| Setup | File | Steps | s/step | Time per clip |
+|---|---|---|---|---|
+| ComfyUI repack, pruned INT8 ConvRot | ? | 20 | 3.87 | 87 s |
+| this work: NVFP4, local-Hessian scales | 12.7 GB | 20 | 3.53 | 79 s |
+| rockerBOO: NVFP4 MLP + INT8 ConvRot attention, edge blocks BF16 (20 GB) | 20.1 GB | 20 | 4.14 | 92 s |
+| INT8 ConvRot + 4-step LoRA (merged) | ? | 4 | – | 24 s |
+| NVFP4 + 4-step LoRA merged (ComfyUI re-quantizes the layers) | 12.7 GB | 4 | – | 23 s |
+| NVFP4 + 4-step LoRA in bypass mode (weights untouched) | 12.7 GB | 4 | – | 26 s |
+| INT8 ConvRot + fast-attention config LoRA | 21.0 GB | 20 | 2.19 | 53 s |
+| this work NVFP4 + fast-attention config LoRA | 12.7 GB | 20 | 1.76 | 43 s |
+| rockerBOO NVFP4 + INT8 ConvRot + fast-attention config LoRA | 20.1 GB | 20 | 2.46 | 58 s |
+<!-- END:local -->
+
+## Files and usage
+
+| File | What | Size |
+|---|---|---|
+| `diffusion_models/minimax_h3_fl2va_pruned_nvfp4_calibrated.safetensors` | FL2VA / T2VA transformer, all block GEMMs NVFP4 (local-Hessian), AdaLN curve form rank 12 | 12.7 GB |
+
+It is a stock ComfyUI "pruned" (curve-form AdaLN) H3 diffusion model: put it in `models/diffusion_models/` and pick it
+in the H3 template's **Load Diffusion Model** node; keep the usual H3 text encoder and VAEs. "Pruned" here means the
+13B AdaLN parameters are replaced by an equivalent timestep table plus rank-12 projections — nothing else is removed.
+
+- **Speed:** NVFP4 GEMMs run natively on Blackwell (RTX 50-series, B200); other GPUs load it but dequantize.
+- **Attention dominates each step on an RTX 5090.** A per-layer attention config LoRA (kitchen INT8 / SOL sparse
+  attention) roughly halves the step time; it carries no weights, so it keeps this file's calibrated scales.
+- **Weight LoRAs** (e.g. 4-step distillation) are merged by ComfyUI by re-quantizing the layers with fresh scales,
+  which discards this file's calibration; ComfyUI's bypass LoRA loader keeps it, at ~15% speed cost.
+
+## How these were made
+
+1. Factor every block's AdaLN through a PCA basis of `silu(temb(t))` (rank 12, fp32).
+2. Record the transformer's exact inputs/outputs at every denoising step for 8 evaluation and 16 calibration
+   prompts.
+3. Calibrate each ModelOpt recipe on the calibration traces; score one-step error on the evaluation traces.
+4. Export to ComfyUI's format, packing weights from the calibrated ModelOpt quantizers (no re-quantization) and
+   checking that the stored tensors decode to exactly what was scored.
+
+## License and attribution
+
+MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights
+Reserved. These files are quantized derivatives of MiniMax-H3 and are distributed under the same agreement,
+including its Acceptable Use Policy and territorial restrictions; see the [license](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE).
+The H3 text encoder is Qwen3-VL-32B (Apache 2.0). Not affiliated with or endorsed by MiniMax or NVIDIA.

assets/compare_chef_talk.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63419c2d893c36ef054b43edd5a03fb505579e221b48f587dae7f8e7e1e30df4
+size 1654896

assets/compare_chef_talk_4step.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a078e43a1a45f7d0489c0e87ab8e768320095666d36a6a8ba474ab6dbe4975d8
+size 1753371

assets/compare_chef_talk_4step_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:634ac639843fa40fc20466fb262a1e667bb811484b8d6688dcae966ca2fdb4bb
+size 5929054

assets/compare_chef_talk_fastattn.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2851da7762d80a9b163d476658bc5fa3a939c2da55b66ca9122ff329a408ad1d
+size 1901586

assets/compare_chef_talk_fastattn_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0383bb5ab857fbf808d70a13a30fadc45d0c1ea25cc6f8e8d4857dab46ec5e4b
+size 6083702

assets/compare_chef_talk_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5014f1c06628ec11eae12cf27fa072252a5f0dd0b0fe7987df5d65ca8085a96
+size 5294808

assets/compare_fox_snow.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23be6e3d5891fb35766c2309c70980d3486c6091b21872e4a4d83511242bb3b1
+size 1556676

assets/compare_fox_snow_4step.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fe22a222be8cf3b7b03e86d217e6cfd61b55744e029074fdf700f20795b793f
+size 3078091

assets/compare_fox_snow_4step_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b197fa085f9af948ed0458565aadcb5d43a92975e511ca5718bc2ff5595852e
+size 11844375

assets/compare_fox_snow_fastattn.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eec25cb4202685480cc018c5ae3c006d45d23061fad75eb582d14e08b2819119
+size 1899872

assets/compare_fox_snow_fastattn_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba5fb0bd277338780b843144c40d2126a7464d84aa1d8705fbecef6c3a5f66f7
+size 6453179

assets/compare_fox_snow_sequential.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0591d5a1d954eb6f2981dcac4a0c29b88d0e0ad7df99ddb0d760882f210a633
+size 5499297