FAQ
June 30, 2026 · View on GitHub
This document collects the questions most frequently raised by the community since SenseNova U1 was open-sourced, organized by topic. If your question isn't covered here, feel free to reach out via GitHub Issues or our Discord community.
Useful links
💡 Note: The Free Token plan for global users will update in July. Stay tuned on Discord!
1. Product & Experience
Q: Where can I request an invite code for the interleaved text-and-image feature?
A: The interleaved text-and-image feature is now available. You can try it using the SenseNova Interleaved demo platform.
Q: Is there a website where I can try it online directly?
A: Yes, you're welcome to use our online demo platform.
Q: Where can I find benchmark comparisons between U1 and other leading models?
A: See Chapter 5 (Evaluation) of the technical report.
Q: Can it generate a 3x3 grid or storyboard? For example, taking a reference image or style as input and producing a 9-panel storyboard.
A: The architecture is capable of supporting this, but the feature itself will come in a future release. Stay tuned.
Q: The infographics generated in the official Studio look quite different from what I get with local deployment or the API. Why is that?
A: The models behind the official Studio and API are tuned for overall cost-effectiveness, balancing efficiency with output quality, which makes them the best all-around fit for most general use cases—and they're updated on a regular basis. The local deployment version offers more flexibility, letting you tailor the setup to your specific needs for more targeted results.
Q: Chinese prompts work better than English ones. Will you optimize for English?
A: Targeted improvements are coming soon—keep an eye out for upcoming releases.
Q: The infographic feature is very useful, but other image types (such as photorealistic portraits) are limited. Will this improve?
A: Infographics are the model's primary focus right now, but we're also continuously improving generation quality for other image types. We welcome your ongoing feedback and suggestions.
2. API & Token Plan
Q: When will the API be released?
A: The Token Plan is already live in mainland China, supporting 1,500 calls per 5-hour window. Token Plans for other countries and regions are scheduled to launch in July, stay tuned.
Q: How do I get an API key to make calls?
A: We recommend creating an API key through the Token Plan. SenseNova 6.7 Flash-Lite is currently free to use—you can sign up at this website (Chinese users only; global version coming in July). See the documentation for details.
Q: Is there a lighter, framework-agnostic way to make U1 directly compatible with the OpenAI API spec?
A: Take a look at SenseTime's Token Plan, which lets you call the API out of the box. Documentation.
3. Technical Report & Paper
Q: When will the technical report be published?
A: It's already out—read it here.
Q: When will the arXiv paper be released?
A: It's already online.
4. Training & Fine-tuning
Q: When will the training code be open-sourced?
A: The full-parameter fine-tuning code for SenseNova-U1 was officially open-sourced on May 21.
Q: Was video data used during training?
A: Video data was used to pre-train the understanding component, though its share in this round of training was relatively small; we expect to increase the ratio in the next version. Video frames are handled roughly as follows: a timestamp is prepended to the image token of each frame—for example, [mm:ss.ff]:<image>. A more detailed description will be provided in a future technical report.
Q: Will the LoRA training code be open-sourced?
A: Open-sourcing the LoRA training code is on our roadmap. Watch for our upcoming releases.
Q: For fixing distorted portraits with full-parameter fine-tuning, roughly how much training data is needed?
A: Issues like this are better addressed by having the model learn them earlier in training; relying on later fine-tuning data alone yields limited returns. Our next model version is focused on solving this problem.
Q: Do visual understanding tasks support multi-image input? I only see single-image examples on HuggingFace.
A: Multi-image input is supported. The underlying NEOChatModel.chat() natively handles multiple images: in samples.jsonl, write the image field as an array of paths, and manually place the corresponding number of 〈image〉 placeholders in the question (we recommend the "Image-1: 〈image〉 newline Image-2: 〈image〉 ..." template). Example: {"id":"cmp","image":["a.jpg","b.jpg"],"question":"Image-1: 〈image〉 newline Image-2: 〈image〉 newline Compare them."} (in actual use, 〈image〉 is the standard image placeholder tag and "newline" is the newline character \n). If you use examples/vqa/inference.py, modify answer() to call load_image_native separately for each image and then torch.cat the pixel_values and grid_hw, while making it accept image as a list—no other changes are needed.
Q: Will the RL (reinforcement learning) code be open-sourced?
A: Open-sourcing the RL code is on our roadmap. Stay tuned.
5. Deployment & Hardware
Q: How do I run it on my local machine? Which deployment framework do you recommend?
A: U1 can be deployed locally. Framework recommendations: ollama (easy to use, but average performance); llama.cpp (slightly more involved to set up, but the best performance); vLLM (fast and GPU-friendly). If you'd rather not deploy it yourself, you can also create an API key through the Token Plan and call it directly (China only; global version coming in July).
Q: How much VRAM is needed to generate a 2K image?
A: Our tests on an H100 showed 35–37 GB of VRAM. See Issue #87 for details.
Q: What resolution gives the best results?
A: We currently recommend 2K, which was also the primary target resolution during internal development.
Q: Can I deploy SenseNova-U1-8B-MoT on a 5090 32G GPU?
A: Yes, it can be deployed. Some users have reported slow API calls and low VRAM utilization; we're looking into this and welcome reports with specific environment details via GitHub Issues to help us pinpoint the cause.
Q: Is there an all-in-one bundle for local deployment?
A: Please follow the quick-start guide to deploy.
Q: Do you support arm64 / aarch64 architectures? Right now both the uv install and the docker image are amd64.
A: The default dependencies are torch 2.8 + cu128, a combination that currently has no corresponding aarch64 wheel. You can try switching to version 2.7.1, or 2.9 and above—PyTorch officially provides some aarch64 wheels: https://download.pytorch.org/whl/cu128/torch/. Some community users have reported successfully installing torch 2.9.0 on aarch64 (with a warning about the version mismatch from the default). We're evaluating whether to switch the default to a version with better aarch64 support.
Q: Is vLLM deployment supported?
A: Yes, vLLM deployment is now supported. See Issue #93.
Q: Is MLX supported?
A: MLX is not supported at this time.
Q: Is NVFP4 quantization supported (for example, a mixed-precision setup on DGX Spark with LightLLM + LightX2V using NVFP4 for text and full precision for vision)?
A: NVFP4 is not supported at this time.
Q: Do you support deployment on Ascend cards / Atlas accelerator cards / Huawei domestic chips?
A: Not at this time—deployment is only supported on Linux + CUDA platforms. Day-0 support for domestic chips is led by the respective chip vendors, and there is no unified official support plan yet.
Q: Is arm64 deployment of lightllm_lightx2v supported?
A: lightllm_lightx2v does not currently support arm64; it only supports Linux + CUDA platforms.
Q: Memory / VRAM usage is high and machines with lower specs can't run it. Is there a more resource-efficient option?
A: The model does have fairly high memory / VRAM requirements. When resources are limited, we recommend using the GGUF low / balanced tiers.
Q: GGUF balanced mode runs on 8 GB, but full mode (e.g. Q6) blows past the VRAM limit. Is that normal?
A: Yes. Full mode has high VRAM requirements, and this is expected. When resources are limited, we recommend switching to the GGUF + VRAM modes low / balanced tiers.
Q: Is there a multi-GPU (2-card / 4-card) inference script? A single card runs out of VRAM.
A: You can use LightLLM + LightX2V for multi-GPU inference, which already supports multi-GPU and multi-node setups. See the deployment guide.
Q: Is there a Python version requirement? Does it have to be 3.11, for instance?
A: The model itself doesn't restrict the Python version. There is, however, a binding between PyTorch and Python versions, so if you can set up the environment yourself, you're free to use your existing Python version.
Q: Why weren't the ComfyUI nodes uploaded to the official ComfyUI node Registry? How do I install them?
A: The official ComfyUI nodes have been published and registered in the ComfyUI node Registry. For installation, see this documentation.
Q: After turning on think_mode, will it run 1024 steps and take far too long?
A: The number of inference steps is capped at either 50 or 8 steps, and this isn't affected by think_mode. If you observe unusual run times in practice, feel free to report your specific environment via GitHub Issues.
Q: When using the infographic feature, 50 steps is actually a few minutes faster than 8 steps. Is that normal?
A: The official Studio, API, and local deployment models have all been upgraded to a new version. We recommend re-testing with the same prompt on the new model. If the issue persists, feel free to report it via GitHub Issues.
Q: On XPU (Intel GPU), it's very slow, and even after setting dtype/device it still ends up running as f32/cpu. What can I do?
A: The XPU-related adaptations have been updated, so we recommend pulling the latest code and trying again. Also, the infographic model may throw a cast error when using f16; switching to bf16 can help. If problems persist, feel free to report them via GitHub Issues.
Q: What hardware is required to run the model?
A: The bf16 model is about 36 GB in size. Running it in full mode requires a machine with more than 36 GB of VRAM; when resources are limited, we recommend using the GGUF low / balanced tiers. For details, see single-card layered offloading.
Q: With two cards totaling 32 GB of VRAM, I get OOM when using --x2i_server_deploy_mode colocate --tp 2. How do I configure it to use system memory?
A: The bf16 model is about 36 GB in size, so it generally requires a machine with more VRAM, or you can use multi-GPU inference instead.
Q: How do I successfully deploy the 38B model? What configuration is required?
A: The 38B model needs about 76 GB of VRAM. We recommend a dual-card 80 GB machine, or a 4-card 32 GB machine (such as the 5090).
Q: Which settings need to be changed to run on a single card? Should both tp and cfg be set to 1?
A: Yes, just set both to 1.
Q: How can I wrap the script as a vLLM / SGLang service?
A: vLLM deployment is now supported. See Issue #93.
6. Prompts
Q: How should I write prompts? The same prompt produces quite different results across platforms.
A: You can refer to the example outputs and prompt structures in the Prompt Gallery. We also recommend enabling thinking mode so the model reasons about your request before generating—this gives more consistent results.
Q: Can prompt enhancement use a local model, or does it have to connect to the API?
A: Local models are supported. See the documentation. For the prompt template, see this file. In addition, SenseNova 6.7 Flash-Lite is currently free and can be used for prompt enhancement, please refer to the SenseNova Token Plan.
7. Common Errors
Q: When running with the lightx2v image, NeoChatTokenizer fails to load / self.conversation_module is None because conversation.py is missing from the model directory. How do I fix this?
A: This has been fixed—please use the latest branch code. See Issue #170 for details, and for deployment, see this documentation.
Q: Conversations work fine inside the image container, but text-to-image throws NoneType object has no attribute get_conv_template. How do I fix this?
A: This has been fixed—please use the latest branch code and refer to the deployment guide.