{"id":13160,"date":"2026-11-17T11:02:10","date_gmt":"2026-11-17T02:02:10","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=13160"},"modified":"2026-11-17T11:11:19","modified_gmt":"2026-11-17T02:11:19","slug":"tencent-releases-hy-nomas","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-hy-nomas\/","title":{"rendered":"Hy-Nomas Video Generation Model: File List"},"content":{"rendered":"<p><em>Sample outputs are available on the <a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas\">model card<\/a>.<\/em><\/p>\n<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\/tencent\/Hy-Nomas\">tencent\/Hy-Nomas<\/a><\/td>\n<\/tr>\n<tr>\n<td>Family guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-hy-nomas-en\/\">Hy-Nomas guide (2 articles)<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher guide<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent: models and licenses<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-21<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/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>Tencent has released &#8220;Hy-Nomas,&#8221; a world generation model capable of generating videos interactively in multi-turn dialogues through text instructions and camera controls. It is a video generation model based on text and camera trajectories that outputs and expands continuous video worlds segment by segment, allowing users to gradually change scenes and viewpoints by inputting text prompts and camera movement trajectories.<\/p>\n<p>The model integrates a segment-autoregressive multimodal Diffusion Transformer (MMDiT), a text encoder, a camera conditioning module, and a PixelUMM video decoder. It features a persistent Key\/Value (KV) cache mechanism that carries over context between generated segments. With a distilled inference configuration, it outputs 720\u00d71280 resolution, 24 FPS video segments in just 8 denoising steps. Furthermore, a low-memory configuration is provided to enable operation on a single NVIDIA GeForce RTX 4090.<\/p>\n<h2>Specifications<\/h2>\n<p>The specifications confirmed from the model card and published configuration files are as follows:<\/p>\n<ul>\n<li><strong>Architecture<\/strong>: Dense Multimodal Diffusion Transformer (dense MMDiT). Adopts a dual-stream structure that processes video latent representations and text tokens.<\/li>\n<li>Number of MMDiT blocks: 54 double-stream blocks (0 single-stream blocks) &#8211; Hidden size: 2,048 &#8211; Attention heads: 16 heads (128 dimensions per head) &#8211; Feed-forward layer intermediate size: 8,192 (4x expansion) &#8211; Positional encoding: 3D RoPE (temporal, height, and width dimensions are <code>[16, 56, 56]<\/code>, theta 256)<\/li>\n<li><strong>Text Encoder<\/strong>: Utilizes the language model component of &#8220;Qwen\/Qwen2.5-VL-7B-Instruct.&#8221; Projects 3,584-dimensional hidden states into the 2,048-dimensional space of the DiT via a 2-layer token refiner.<\/li>\n<li><strong>Camera Conditioning<\/strong>: Groups normalized 4&#215;4 extrinsic parameter matrices in units of 4 frames and projects them as 64-dimensional inputs. Reflects camera viewpoints and motion in video tokens via camera cross-attention within the double-stream blocks.<\/li>\n<li><strong>Video Latent Space and Decoder<\/strong>: DCAE-V <code>f32t4c32<\/code> latent representation (32 channels, 32x compression spatially, 4x compression temporally). Output latent representations are restored using a fine-tuned PixelUMM decoder (36 layers, hidden 4,096, query heads 32, KV heads 8).<\/li>\n<li><strong>Output Specifications`<\/strong>:<\/li>\n<li>Default resolution: 720 \u00d7 1280 (480 \u00d7 832 and others can also be specified by changing settings) &#8211; Frame rate: 24 FPS<\/li>\n<li><strong>Recommended Inference Settings`<\/strong>:<\/li>\n<li>DiT denoising steps: 8 steps (when using the distilled model) &#8211; PixelUMM decoding steps: 1 step &#8211; Solver: Euler ODE (flow velocity prediction) &#8211; Guidance scale (CFG): Text 1.0, Camera 1.0 &#8211; Camera movement\/rotation intensity: Translation scale 1.0, Rotation scale 2.5<\/li>\n<li><strong>License<\/strong>: Apache-2.0 license is applied. It is an open license system that allows commercial use, modification, and redistribution.<\/li>\n<\/ul>\n<h2>Performance and Quality<\/h2>\n<p>While benchmark test score numerical tables are not directly posted in the model card by the publishers, &#8220;Hy-Nomas-Bench&#8221; has been simultaneously released as an evaluation suite to measure interactive controllability and generation consistency. Hy-Nomas-Bench consists of a total of 280 test cases covering four areas: multi-turn free-form text dialogue, camera navigation, interleaved text and camera control, and spatial memory, designed to comprehensively verify continuous world generation performance.<\/p>\n<p>Additionally, in terms of quality and inference efficiency, the publishers indicate that the introduction of distillation technology enables the generation of high-quality 720p 24 FPS videos with just 8 steps of denoising and 1 step of PixelUMM decoding. Furthermore, adapters are provided such as &#8220;Cinematic LoRA&#8221; to enhance visual quality and &#8220;Robotics LoRA&#8221;\u2014trained on a real-world robotic arm dataset (RealSource-World)\u2014to enhance robotics simulation performance in world generation, allowing the expansion of expressive power according to the application.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Hy-Nomas excels at interactively constructing and expanding dynamic virtual worlds by combining free-form text instructions with camera movement trajectories. Unlike conventional single-shot video generation models, it can retain past context as a KV cache while generating videos segment by segment, smoothly introducing new events and viewpoint changes while maintaining the generated worldview and spatial consistency.<\/p>\n<p>Specifically, the following uses and scenarios are anticipated:<\/p>\n<ul>\n<li><strong>World Building via Multi-Turn Dialogue<\/strong>: After generating an initial scene, providing additional text instructions such as &#8220;a fox passes through the flowers&#8221; or &#8220;the weather changes to night and glowing flowers bloom&#8221; allows the subject&#8217;s actions, appearance, overall scene style, and environmental conditions to be changed step by step.<\/li>\n<li><strong>Precise Camera Navigation<\/strong>: By inputting 3D camera trajectories (translation and rotation) defined by extrinsic parameter matrices, users can create footage with meticulous camera work matching directorial intent, such as first-person or third-person movement and spatial exploration from free angles.<\/li>\n<li><strong>Robotics Simulation<\/strong>: By integrating &#8220;Robotics LoRA&#8221;\u2014trained on the real-world robotic arm dataset (RealSource-World)\u2014into the distilled model, it can be utilized for generating simulation footage for robot motion verification and environment recognition.<\/li>\n<li><strong>Cinematic Video Production<\/strong>: By utilizing the provided &#8220;Cinematic LoRA,&#8221; users can easily pursue cinematic visual expressions with high-quality lighting and composition.<\/li>\n<li><strong>Interactive Demos and Prototyping<\/strong>: Because it can reflect instructions with a nearly real-time feel, it is also suitable for scene generation and exploratory content verification in game development and virtual space prototyping.<\/li>\n<\/ul>\n<p><!-- lmw:files --><\/p>\n<h2>Distributed Files<\/h2>\n<p><em>Weight files published in <a href=\"https:\/\/huggingface.co\/tencent\/Hy-Nomas\/tree\/main\">tencent\/Hy-Nomas<\/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>DiT\/base_ar\/diffusion_pytorch_model_fp32.safetensors<\/code><\/td>\n<td>34.21GB<\/td>\n<\/tr>\n<tr>\n<td><code>DiT\/distilled\/diffusion_pytorch_model_ema_0_995.safetensors<\/code><\/td>\n<td>34.21GB<\/td>\n<\/tr>\n<tr>\n<td><code>DiT\/distilled_int8\/diffusion_pytorch_model_ema_0_995_int8.safetensors<\/code><\/td>\n<td>9.06GB<\/td>\n<\/tr>\n<tr>\n<td><code>DiT\/lora_weights\/cinematic\/adapter.safetensors<\/code><\/td>\n<td>6.12GB<\/td>\n<\/tr>\n<tr>\n<td><code>DiT\/lora_weights\/dual_arm_robotics\/adapter.safetensors<\/code><\/td>\n<td>12.24GB<\/td>\n<\/tr>\n<tr>\n<td><code>PE\/model-*-of-00018.safetensors<\/code><\/td>\n<td>55.56GB (18 split files)<\/td>\n<\/tr>\n<tr>\n<td><code>VAE\/DCAE-V\/dc-ae-v-1.0-f32t4c32-bf16.pt<\/code><\/td>\n<td>4.28GB<\/td>\n<\/tr>\n<tr>\n<td><code>VAE\/PixelUMM_FT\/diffusion_pytorch_model.safetensors<\/code><\/td>\n<td>31.55GB<\/td>\n<\/tr>\n<tr>\n<td><code>VAE\/PixelUMM_FT_FP8\/diffusion_pytorch_model_fp8.safetensors<\/code><\/td>\n<td>9.13GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:files --><\/p>\n<h2>How to Get It<\/h2>\n<p>Hy-Nomas model weights are distributed in the Hugging Face <code>tencent\/Hy-Nomas<\/code> repository. Since it is not a gated model, prior agreement to terms of use is not required, and anyone can directly download and use it.<\/p>\n<p>Model weights are mainly provided in safetensors format, and appropriate files must be gathered according to the application and execution environment.<\/p>\n<ul>\n<li><strong>DiT Model<\/strong>: In addition to the distilled version (<code>DiT\/distilled\/diffusion_pytorch_model_ema_0_995.safetensors<\/code>) used for standard 8-step inference and web demos, a quantized version for single-GPU environments (<code>DiT\/distilled_int8\/diffusion_pytorch_model_ema_0_995_int8.safetensors<\/code>) and a non-distilled base model (<code>DiT\/base_ar\/diffusion_pytorch_model_fp32.safetensors<\/code>) are prepared.<\/li>\n<li><strong>Decoder<\/strong>: The standard PixelUMM decoder (<code>VAE\/PixelUMM_FT\/diffusion_pytorch_model.safetensors<\/code>) and an FP8 quantized version for low-memory configurations (<code>VAE\/PixelUMM_FT_FP8\/diffusion_pytorch_model_fp8.safetensors<\/code>) are distributed.<\/li>\n<li><strong>Adapters<\/strong>: <code>DiT\/lora_weights\/cinematic\/<\/code> for image quality improvement and <code>DiT\/lora_weights\/dual_arm_robotics\/<\/code> for robotics applications are also available from the same repository.<\/li>\n<\/ul>\n<p>Inference code is published in the official GitHub repository (<a href=\"https:\/\/github.com\/Tencent-Hunyuan\/Hy-Nomas\">https:\/\/github.com\/Tencent-Hunyuan\/Hy-Nomas<\/a>). As for the execution environment, instructions are provided to use the official recommended Docker image (<code>hynomasoss\/hynomasoss:cu130<\/code>), which collectively installs the Python environment, PyTorch, and necessary CUDA attention kernels.<\/p>\n<p>When setting up directly in a local environment, after preparing Python 3.12 or higher, uv, and a CUDA 13.0 Toolkit environment, dependencies are installed from the repository root.<\/p>\n<pre><code class=\"language-bash\">git clone https:\/\/github.com\/Tencent-Hunyuan\/Hy-Nomas.git\ncd Hy-Nomas\nuv sync --extra all\nsource.venv\/bin\/activate\n<\/code><\/pre>\n<p>The weights for <code>Qwen\/Qwen2.5-VL-7B-Instruct<\/code> used as the text encoder are automatically fetched from Hugging Face when inference is executed (local paths can also be specified directly).<\/p>\n<p>Inference is executed via shell scripts. When running with the standard configuration (8-way sequence parallelism), run the following command:<\/p>\n<pre><code class=\"language-bash\">export HY_NOMAS_MODEL_DIR=&quot;\/path\/to\/Hy-Nomas-models&quot;\n\nbash scripts\/inference\/inference_from_config.sh configs\/experiment\/inference.yaml \\\n  dit_weight=&quot;$HY_NOMAS_MODEL_DIR\/DiT\/distilled\/diffusion_pytorch_model_ema_0_995.safetensors&quot; \\\n  model.vae.checkpoint=&quot;$HY_NOMAS_MODEL_DIR\/VAE\/PixelUMM_FT\/diffusion_pytorch_model.safetensors&quot; \\\n  inference.prompt=&quot;$HY_NOMAS_MODEL_DIR\/examples\/prompts.json&quot; \\\n  inference.save_path=.\/outputs\/distill\n<\/code><\/pre>\n<p>When running on a low-memory environment such as a single GeForce RTX 4090, specify the INT8 DiT and FP8 PixelUMM, and execute using the dedicated configuration file.<\/p>\n<pre><code class=\"language-bash\">bash scripts\/inference\/inference_from_config.sh configs\/experiment\/single_gpu_low_mem_inference.yaml \\\n  dit_weight=&quot;$HY_NOMAS_MODEL_DIR\/DiT\/distilled_int8\/diffusion_pytorch_model_ema_0_995_int8.safetensors&quot; \\\n  model.vae.checkpoint=&quot;$HY_NOMAS_MODEL_DIR\/VAE\/PixelUMM_FT_FP8\/diffusion_pytorch_model_fp8.safetensors&quot; \\\n  inference.prompt=&quot;$HY_NOMAS_MODEL_DIR\/examples\/prompts.json&quot; \\\n  inference.save_path=.\/outputs\/single_gpu\n<\/code><\/pre>\n<p>Additionally, an online interactive demo operable from a web browser is also published, making it possible to try out operation and generation quality before preparing equipment on hand.<\/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>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/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:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/hy-nomas-diffusers-released\/\">Hy-Nomas-Diffusers Interactive Video Generation Model: File List<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-qwen3-8-27b-dspine-b16\/\">Qwen3.8-27B-DSpine-b16 Draft Model for Speculative Decoding: 8GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-wevisone-vl-vision-language-models\/\">Tencent Releases WeVisOne-VL Vision Language Models<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/tencent-releases-dspine-parallel-draft-model\/\">Qwen3-8B-DSpine-b16: 4GB+ VRAM<\/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-tencent-hy-nomas-en\/\">Hy-Nomas family overview (2 articles)<\/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-tencent-en\/\">Tencent: 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\/tencent\/Hy-Nomas\">tencent\/Hy-Nomas (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/Tencent-Hunyuan\/Hy-Nomas\">Tencent-Hunyuan\/Hy-Nomas (GitHub)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/datasets\/tencent\/Hy-Nomas-Bench\">tencent\/Hy-Nomas-Bench (Hugging Face Datasets)<\/a><\/li>\n<li><a href=\"https:\/\/hy-nomas.hy.tencent.com\/\">Hy-Nomas Interactive Demo<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Explore Hy-Nomas, Tencent&#8217;s world generation model enabling multi-turn interactive video generation via text prompts and camera control.<\/p>\n","protected":false},"author":1,"featured_media":13159,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[203],"tags":[4060,4062,1172,1547,3240,959,1746],"class_list":["post-13160","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-image-video-and-audio","tag-hy-nomas-en","tag-mmdit-en","tag-tencent-en","tag-verified","tag-video-generation-en","tag--en"],"lang":"en","translations":{"en":13160,"ja":13158},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13160","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=13160"}],"version-history":[{"count":3,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13160\/revisions"}],"predecessor-version":[{"id":17923,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13160\/revisions\/17923"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/13159"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=13160"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=13160"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=13160"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}