SenseNova U1 open weights released

SenseTime’s SenseNova team has now put enough material in public view to make SenseNova U1 more than a teaser release. The project includes public code on GitHub, model checkpoints on Hugging Face, and a technical report on arXiv describing the architecture, training setup and benchmark results. The initial 8B model weights were released on April 27, according to the project repository, and the team said on May 10 that it had added the technical report and weights for the larger A3B variant. The release matters because SenseNova U1 is being presented as a single multimodal system rather than a stitched pipeline. The paper says the family is built to handle understanding, reasoning and generation in one architecture, and the company’s April 29 announcement described it as a “native unified multimodal model series.” ### So what, exactly, is open here? GitHub and Hugging Face pages show public access to the SenseNova-U1-8B-MoT and SenseNova-U1-A3B-MoT checkpoints, along with supervised fine-tuned versions of both. (github.com) The repository is published under an Apache-2.0 license, and the Hugging Face collection lists the 8B and A3B model entries as updated within the past several days. (arxiv.org) The repository changelog gives the clearest timeline. April 27 is listed as the initial release date for the 8B model and inference code, while May 10 is listed as the date for the technical report and the A3B-MoT and A3B-MoT-SFT weights. ### Why do some posts call it 38B, while the paper says 30B-A3B and Hugging Face shows 39B? The paper describes two variants: SenseNova-U1-8B-MoT and SenseNova-U1-A3B-MoT, with the larger system built on a “mixture-of-experts (30B-A3B) understanding baseline.” Hugging Face, meanwhile, labels the A3B entries as “39B,” while third-party posts have described the model as roughly 38B. (github.com) That leaves a naming mismatch in public materials rather than a clean single figure. The most defensible reading from the primary sources is that “A3B” refers to about 3 billion activated parameters in a mixture-of-experts setup, while total parameter counts are presented differently across the paper and hosting pages. Neither the GitHub summary nor the Hugging Face collection page shown in search results resolves that discrepancy on its face. (arxiv.org) ### What does SenseNova say is different about the architecture? SenseTime said on April 29 that SenseNova U1 is built on its NEO-unify architecture and “completely eliminates” both the visual encoder and the variational auto-encoder used in many multimodal systems. The company said the goal is a unified representation space in which images and text are processed in the same framework rather than handed between separate components. (arxiv.org) The arXiv paper makes the same case in research language. It says current vision-language systems are often split between understanding and generation, and that SenseNova U1 is designed so those functions “evolve as synergistic views of a single underlying process.” ### What can the public models actually do? The paper says the models target both understanding tasks and generation tasks, including text understanding, visual perception, reasoning, any-to-image synthesis, infographic generation and interleaved image-text generation. (sensetime.com) The company announcement highlighted continuous image-text creation and infographic generation as flagship use cases. (arxiv.org) The public project pages also show the release is still moving. GitHub updates on May 8 added GGUF quantized checkpoints and low-VRAM inference options, and a May 15 Hugging Face update added a separate infographic-focused model. ### Where should people look if they want the primary materials? The most useful starting points are the GitHub repository, the Hugging Face SenseNova-U1 collection and the arXiv report. (arxiv.org) Those three sources together provide the release timeline, model files and the team’s own technical description of the 8B and A3B variants. As of May 17, 2026, the repository remains active, the Hugging Face collection shows recent updates, and the technical report submitted on May 12 is the main public document for architecture and training details. (github.com)