{"id":12621,"date":"2026-11-17T11:01:29","date_gmt":"2026-11-17T02:01:29","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=12621"},"modified":"2026-11-17T16:11:36","modified_gmt":"2026-11-17T07:11:36","slug":"fasth3-qad-v2-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-qad-v2-released\/","title":{"rendered":"FastH3-QAD-v2 Audio-Enabled Video Generation Model: File List"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-QAD-v2\">FastVideo\/FastH3-QAD-v2<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-minimaxai-minimax-h3-en\/\">MiniMax-H3 guide (14 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-minimax-en\/\">MiniMax: models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-17<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>other<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">safetensors<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>FastVideo has released &#8220;FastH3-QAD-v2&#8221;, an audio-video generation model with sound. By inputting text, it outputs video accompanied by synchronized audio.<\/p>\n<p>This model is a derivative model quantized to the NVFP4 format, based on the audio-video integrated model &#8220;FastVideo-FastH3-8-Step-V2&#8221;, which supports 8-step inference. It is trained using Quantization-Aware Distillation (QAD) and supports execution on a single Blackwell-generation GPU environment (such as the RTX 5090 or RTX PRO 6000) or DGX Spark.<\/p>\n<h2>Specifications<\/h2>\n<ul>\n<li><strong>Base Model<\/strong>: Based on &#8220;FastVideo\/FastVideo-FastH3-8-Step-V2&#8221; (the original root base model is MiniMaxAI\/MiniMax-H3).<\/li>\n<li><strong>Transformer Architecture<\/strong>: Employs a 50-block Transformer configuration. NVFP4 W4A4 is applied to all attention layers (q, k, v, out), MLP linear layers, and sparse attention (VSA) gates. While the weights for q\/k\/v\/out and MLP are trained via QAD, the VSA gates and all normalization layers (Norm) are kept in BF16.<\/li>\n<li><strong>Attention Mechanism<\/strong>: Employs block-sparse VSA with 64-token tiles (sparsity of 0.8). Although the checkpoints themselves are trained with BF16 VSA, FastVideo also provides an FP4 VSA inference lane for Blackwell.<\/li>\n<li><strong>Text Encoder<\/strong>: Uses Qwen3-VL in NVFP4 format, trimmed to match the 50 layers read by MiniMax H3.<\/li>\n<li><strong>VAE<\/strong>: Uses a lightweight video VAE &#8220;LynnReal&#8221; with INT8 weights applied by Kijai on the video side, combined with the H3 audio VAE on the audio side.<\/li>\n<li><strong>Inference Steps and Sampling<\/strong>: Adopts an 8-step DMD (Distribution Matching Distillation) schedule. Settings are recorded in the bundled <code>fastvideo_inference.json<\/code> and automatically loaded by FastVideo.<\/li>\n<li><strong>License Terms<\/strong>: Set as an &#8220;other&#8221; license on the repository. Additionally, the original model FastVideo-FastH3-8-Step-V2 notes that it inherits the &#8220;MiniMax H3 Community License&#8221;.<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>Measurement results for the prediction error against the BF16 version of the teacher model (relative L2 error of x0 prediction using 30 holdout prompts) have been published by the release source. This metric indicates that lower values are superior.<\/p>\n<p>The measurement results stated in the model card are as follows:<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Metric<\/th>\n<th style=\"text-align: right;\">FastH3-QAD-v2<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Mean (video + audio)<\/td>\n<td style=\"text-align: right;\">0.351<\/td>\n<\/tr>\n<tr>\n<td>Video<\/td>\n<td style=\"text-align: right;\">0.382<\/td>\n<\/tr>\n<tr>\n<td>Audio<\/td>\n<td style=\"text-align: right;\">0.319<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to the published figures, the mean relative L2 error (Mean) combining video and audio is 0.351. Looking at the breakdown, the relative L2 error for video alone is 0.382, whereas for audio alone it is 0.319, indicating that the error relative to the teacher model is kept smaller on the audio generation side.<\/p>\n<p>Regarding QAD, the quantization method used, the release source explains that the NVFP4 student model was trained along the 8-step trajectory of the frozen BF16 teacher model. Because the student model executes forward passes using actual 4-bit weights and activations, it acquires weights suitable for quantized computation. Training was executed for 300 steps using only text prompts with a learning rate of 5e-6, utilizing the student&#8217;s own trajectory for 25% of the rows. Because modifying VSA gates or normalization layers changes not only the attention calculation precision but also the video tiles themselves that are attended to, these are frozen, and only the linear layers targeted for quantization are trained.<\/p>\n<p>Regarding the performance aspects of the original model FastVideo-FastH3-8-Step-V2, while it can generate video and audio in just 8 Transformer forwards using VSA-H3 with 80% sparsity and data-free DMD2, FL2VA and Ref2VA are said not to be distilled. Furthermore, the original model card notes that complex motion, fine details, and certain audio qualities may be inferior to the base MiniMax H3 model.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>This model specializes in text-to-video generation with audio. Its standout feature is the ability to simultaneously output video and synchronized audio in an extremely low number of steps\u2014just 8 forward passes.<\/p>\n<p>Anticipated use cases include the generation of short-form video content with audio and prototyping. It primarily targets local inference on a single machine, such as a single Blackwell-generation GPU environment (like the RTX 5090 or RTX PRO 6000) or DGX Spark. It is well-suited for building operational environments that can efficiently execute processes on a single GPU that previously required multi-GPU configurations in base models.<\/p>\n<p>Moreover, rather than simply quantizing existing weights via post-processing, it learns weights suitable for 4-bit forward passes through Quantization-Aware Distillation (QAD), making it ideal for scenarios where high-quality synchronized video is desired while suppressing errors associated with quantization. However, as noted in the original model information, users must keep in mind that it does not support tasks such as FL2VA and Ref2VA, and complex motions, detailed descriptions, and certain audio qualities may be subject to limitations from the base MiniMax H3 model.<\/p>\n<h2>How It Differs from Similar Models<\/h2>\n<p>The models introduced in our site&#8217;s previous coverage, <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/06\/fasth3-v2-fp8-nvfp4-released\/\">FP8 and NVFP4 Versions of Audio-Video Generation Model &#8220;FastH3 V2&#8221; Released<\/a>, performed quantization by inheriting calibrated weights and activation scales. In contrast, the current FastH3-QAD-v2 differs by adopting QAD (Quantization-Aware Distillation), which directly distills and trains an NVFP4 student model along the generation trajectory of a frozen BF16 teacher model, taking an approach that further suppresses errors during 4-bit execution.<\/p>\n<p>Additionally, compared to <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/06\/fasth3-trim-8-step\/\">8-Step Version of Audio-Video Generation Model &#8220;FastH3 Trim&#8221; Released<\/a>, which removed 8 Transformer layers to reduce the architecture to a 42-block configuration, FastH3-QAD-v2 maintains the original 50-block configuration entirely. Rather than structural simplification through block reduction, it aims to balance single-GPU executability and output precision by combining NVFP4 W4A4 quantization and distillation while preserving the scale of the architecture.<\/p>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-QAD-v2\/tree\/main\">FastVideo\/FastH3-QAD-v2<\/a>, listed by this site from the Hugging Face API. Sizes are the actual file sizes.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>File<\/th>\n<th>Size<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>audio_vae\/diffusion_pytorch_model.safetensors<\/code><\/td>\n<td>605MB<\/td>\n<\/tr>\n<tr>\n<td><code>text_encoder\/model-*-of-00004.safetensors<\/code><\/td>\n<td>16.46GB (4 split files)<\/td>\n<\/tr>\n<tr>\n<td><code>transformer\/adaln_tables.pt<\/code><\/td>\n<td>155MB<\/td>\n<\/tr>\n<tr>\n<td><code>transformer\/diffusion_pytorch_model-*-of-00006.safetensors<\/code><\/td>\n<td>27.71GB (6 split files)<\/td>\n<\/tr>\n<tr>\n<td><code>transformer\/nvfp4_weights.safetensors<\/code><\/td>\n<td>11.92GB<\/td>\n<\/tr>\n<tr>\n<td><code>vae\/diffusion_pytorch_model.safetensors<\/code><\/td>\n<td>4.23GB<\/td>\n<\/tr>\n<tr>\n<td><code>vae\/minimax_h3_video_vae_int8_convrot.safetensors<\/code><\/td>\n<td>2.14GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>FastH3-QAD-v2 is available in safetensors format from the Hugging Face repository (FastVideo\/FastH3-QAD-v2). The repository is not gated, allowing users to download and use it without agreeing to additional terms of use. It supports the Diffusers library&#8217;s modular pipeline (<code>MiniMaxH3ModularPipeline<\/code>) and is available via the FastVideo framework.<\/p>\n<p>For environments with less than 64GB of system RAM, mechanisms to suppress memory consumption are provided. The AdaLN timestep projection layers for all 50 blocks of this model consume approximately 26GB of capacity when loaded in BF16 format. In response, the release source provides <code>transformer\/adaln_tables.pt<\/code> (155MB), which is pre-computed output data tailored for the bundled 8-step schedule (<code>fastvideo_inference.json<\/code>).<\/p>\n<p>By acquiring this table file in advance and setting it in the environment variables, you can skip loading the 26GB projection layers on the FastVideo side and significantly reduce system RAM consumption. The configuration procedure is as follows:<\/p>\n<pre><code class=\"language-python\">import os\nfrom huggingface_hub import hf_hub_download\n\nos.environ[&quot;FASTVIDEO_H3_ADALN_TABLE&quot;] = hf_hub_download(\n    &quot;FastVideo\/FastH3-QAD-v2&quot;, &quot;transformer\/adaln_tables.pt&quot;)\n<\/code><\/pre>\n<p>Note that since this table is dedicated to the bundled inference schedule, specifying other timesteps will result in an error on the FastVideo side.<\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>other<\/code>:<\/strong> A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:lab --><\/p>\n<h2>Our Own Measurements<\/h2>\n<h3>Measurements We Did Not Take<\/h3>\n<p>We have not confirmed that this site meets the commercial-use terms of this model&#8217;s license (<code>other<\/code>). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).<\/p>\n<p><!-- \/lmw:lab --><\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omni-nvfp4-qad-comfy-released\/\">FastH3-Omni-NVFP4-QAD-Comfy: 16GB+ VRAM, File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/06\/fasth3-v2-quantized-models-released\/\">Two Quantized Versions of Audio-Visual Video Generation Model FastH<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/06\/fasth3-trim-8-step-video-generation-model\/\">FastH3 Trim: 8-Step Text-to-Video Model with Audio Support<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/fasth3-omni-comfy-released\/\">FastH3-Omni-Comfy: 80GB+ VRAM, File List<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Explore the same model family<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-minimaxai-minimax-h3-en\/\">MiniMax-H3 family overview (14 articles, 4 converted builds)<\/a><\/li>\n<li><strong>Formats this model is available in<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-safetensors-en\/\">Safetensors format guide and models<\/a><\/li>\n<li><strong>Learn about the publisher<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-minimax-en\/\">MiniMax: models, licenses and articles<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-video\">Other video generation models<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/FastVideo\/FastH3-QAD-v2\">https:\/\/huggingface.co\/FastVideo\/FastH3-QAD-v2<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/FastVideo\/FastVideo-FastH3-8-Step-V2\">https:\/\/huggingface.co\/FastVideo\/FastVideo-FastH3-8-Step-V2<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>FastH3-QAD-v2 is an audio-video generation model utilizing NVFP4 quantization and QAD, enabling 8-step inference on single Blackwell GPUs.<\/p>\n","protected":false},"author":1,"featured_media":12620,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[3796,3031,730,813,1547,959,2663,213],"class_list":["post-12621","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-fasth3-qad-v2-en","tag-fastvideo-en","tag-minimax-h3-en","tag-nvfp4-en","tag-verified","tag--en"],"lang":"en","translations":{"en":12621,"ja":12619},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12621","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=12621"}],"version-history":[{"count":3,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12621\/revisions"}],"predecessor-version":[{"id":18245,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12621\/revisions\/18245"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/12620"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=12621"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=12621"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=12621"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}