SenseTime Releases SenseNova U1 Pro: Multimodal Model Competition Shifts Toward Complex Task Delivery

1. What happened

On July 18, at the 2026 World Artificial Intelligence Conference (WAIC), SenseTime officially launched the flagship model in its large-model family, SenseNova U1 Pro. SenseTime positions it as a “delivery-grade” native multimodal agent foundation model for long-horizon tasks, aiming to move the model from one-shot content generation to understanding, planning, execution, checking, and correction in complex tasks.

SenseNova U1 Pro belongs to the flagship version of the “日日新 SenseNova U” series, with a focus on scenarios such as visual design, infographics, presentation decks, video storyboards, and scientific illustrations. SenseTime said the preview version is already open for invitation-only testing, while the full version is expected to be made available to the public in August 2026, along with pricing and API services.

2. Why it matters

From generating content to delivering results

Traditional generative models usually revolve around generating an image or other content from a single prompt, while SenseTime wants U1 Pro to handle longer task chains: the model needs to understand complex objectives, continue with multiple rounds of creation and revision, and ultimately deliver a relatively complete visual result.

SenseTime summarizes this direction as moving from “Generation” to “Delivery,” and compares it with the evolution in the coding domain from Copilot, Vibe Coding, to Agentic Coding. If this product direction can be stably implemented, competition in visual models will no longer be determined only by the image quality of a single picture, but also by information organization, stylistic consistency, detail modification, and task completion capability.

Visual models are starting to enter production workflows

U1 Pro is not aimed at a single entertainment-oriented image generation use case, but at workflows such as enterprise infographics and PPTs, commercial posters, e-commerce product visuals, scientific illustrations, film and television concept design, and anime storyboards. Its product positioning shows that SenseTime is trying to further embed multimodal models into office, commercial design, and content production processes.

However, the publicly available information currently comes mainly from SenseTime releases and media reports, which is not enough to independently prove the model’s stability and efficiency across different real-world workflows.

3. Public information and product capabilities

According to SenseTime, SenseNova U1 Pro mainly has the following capabilities:

• Professional visual design: Emphasizes composition, color, and layout effects, aiming to reduce the “AI feel” common in traditional generative images.
• Native 8K output: Supports up to native 8K resolution, targeting the creation of ultra-long, ultra-large, and high-information-density visual content.
• Image-text and detail control: Strengthens handling of relationships among text, images, layouts, and details, suitable for infographics, presentation decks, and complex knowledge illustrations.
• Long-horizon Agentic loop: Supports dozens of agentic generation cycles around complex goals, and supports synchronized editing of overall style and local text.

Examples shown by SenseTime include first planning the worldview, character settings, clothing and weapons, and environmental tone of a video plot, then generating up to 22 logically consistent consecutive storyboards; and integrating advancement paths, player matchups, tactical systems, passing networks, and touch heatmaps from sports games into a high-definition chart report.

In addition, SenseTime also showcased a 4:1 ultra-wide piece containing city landmarks and high-density text information, used to illustrate the model’s capabilities in long canvases, complex layouts, and image-text integration.

4. Technical approach and evidence boundaries

SenseTime says the underlying architecture of U1 Pro adopts the NEO-unify kernel architecture, attempting to unify language and visual representations within a native multimodal system to reduce the gap between understanding and generation.

SenseTime also said that its open-source SenseNova-Vision unified vision large model has already incorporated traditional vision tasks such as instance segmentation and object detection into general large-model capabilities, and the company hopes to strengthen multimodal agents’ understanding of the visual world through this approach.

At present, the public materials have not disclosed U1 Pro’s parameter scale, training data, inference cost, context length, or a complete technical report. Capabilities such as native 8K, low text error rates, and dozens of agentic cycles mainly come from SenseTime releases and media paraphrases, and still require further verification through public examples, third-party benchmark testing, and real user feedback.

5. What to watch next

1. Whether the full version opens as scheduled: The preview version is already open for invitation-only testing, and the full version is expected to launch in August 2026. The actual release date and scope of access still need to be observed.
2. API and pricing: SenseTime plans to launch the corresponding API service, and the billing method, usage limits, inference speed, and 8K output costs will directly affect enterprise adoption.
3. Third-party evaluation results: Independent performance in text accuracy, complex layout, long-horizon task completion, stylistic consistency, and multi-round revisions still needs to be verified.
4. Differences from other products in the same series: The differences among U1 Pro, SenseNova U1, and U1-Fast in capability, speed, price, and applicable scenarios may affect user choice.
5. Commercialization progress: SenseTime says the relevant capabilities have already been validated in products such as “小浣熊” and “Seko,” but it has not yet disclosed specific usage scale, customer retention, or revenue contribution.

Conclusion

The launch of SenseNova U1 Pro reflects how competition in multimodal models is shifting from the generation quality of a single image to the planning, iteration, and delivery of complex tasks. SenseTime is trying to use native 8K, image-text control, and long-horizon agentic loops to move visual models from creative tools to a more complete productivity system.

Whether this positioning holds will ultimately depend on the availability of the full version, API pricing, third-party evaluations, and real user feedback.

Original source

36氪