SenseNova is the family of AI models built by SenseTime, a Chinese AI software company founded in 2014 and listed on The Stock Exchange of Hong Kong. SenseTime unveiled the SenseNova foundation model series on April 10, 2023, and as of November 2026 the name covers SenseNova 6.8 Flash, which SenseTime calls a "multimodal agent model" for office work, plus several openly licensed models published on Hugging Face.
What is SenseTime?
SenseTime describes itself as an AI software company. Its company page says the business spans Generative AI, Vision AI and emerging segments.
Two products sit underneath the models. SenseCore is SenseTime's AI infrastructure, the compute layer used to train and run models. SenseFoundry is its vision AI platform. The same page states that SenseTime "has developed the SenseNova foundation model series".
As of November 9, 2026, SenseTime's site lists 40,400 PFlops of total SenseCore compute capacity. That is a company figure about its hardware pool, not a measure of any single model.
The company is listed on The Stock Exchange of Hong Kong, with the ticker 0020.HK given in a SenseTime release dated November 5, 2026. SenseTime has offices in Hong Kong, Shanghai, Beijing and other cities, and outside Greater China in Singapore, South Korea, Malaysia, Abu Dhabi, Dubai and Riyadh.
SenseTime's company page credits its vision platform with "years of accumulated computer vision expertise", while the SenseNova series and its SenseChat language model arrived in 2023. Image, video and 3D work appear alongside text in SenseNova releases.
When did SenseNova launch?
SenseTime unveiled SenseNova on April 10, 2023, at a Tech Day event in Shanghai. The company's release describes the "SenseNova" foundation model set.
The launch introduced SenseChat, a large language model that SenseTime described as having "hundreds of billions of parameters" and trained with Chinese context in mind. The release quotes Xu Li, chairman and chief executive, and Wang Xiaogang, co-founder and chief scientist.
The 2023 event was not limited to text. SenseTime showed SenseMirage for text-to-image generation, SenseAvatar for digital humans built from a 5-minute video clip, and SenseSpace and SenseThings for 3D content.
On hardware, SenseTime said SenseCore had 27,000 GPUs and 5,000 petaflops at the time of the launch.
What does SenseNova mean?
Xu Li said the name implies "constant renewal, daily renewal, and further renewal". The Chinese name, which SenseTime uses in its Chinese releases, is Ri Ri Xin (日日新), literally "day day new".
In practice SenseNova works as a series name rather than the name of one model. SenseTime uses it across language and multimodal models, image generation products, and openly licensed models published on Hugging Face. Individual releases carry version numbers or suffixes, such as SenseNova 6.8 Flash, SenseNova-U1.5-8B-MoT or SenseNova-6.8-Code2Video.
Because the series name is reused, "SenseNova" by itself does not tell you which model, license or capability set is being discussed. The full release name matters, especially when comparing an API-only model with a downloadable one.

What can SenseNova 6.8 Flash do?
SenseTime released the official version of SenseNova 6.8 Flash in November 2026. SenseTime's release is dated November 5, 2026 in Hong Kong. TechNode reports that the official version was released on November 2. SenseTime calls it a "multimodal agent model" designed for office work and complex real-world tasks.
Two capabilities get the most attention in the release: long-term memory and proactive interaction. SenseTime says the model can recall earlier work context and suggest follow-ups when information changes. The demos given are a sales follow-up after a product upgrade, and an employee handover that uses authorized work history.
The second area is design and document output. SenseTime lists dynamic PPTX presentations, animated posters, web pages, SVG graphics and animation, interactive 3D web pages built with Three.js, and 3D modeling with Blender. The release also mentions integration with Figma. For the Blender 3D modeling feature, SenseTime adds a caveat: "These functions require local environment configuration and multiple rounds of user interaction".
On performance, SenseTime says the model "outperformed some leading international models in multiple benchmark evaluations". Charts in the release name the tests and the rival models; all scores are SenseTime's own. On six agent and office tests, SenseNova 6.8 Flash has the top score shown on TeamBench (53.33) and DataAnalysis2.0 (86.89), and trails a rival on the other four: AutomationBench (49.89 vs 50.30 for Opus 5), WorkbuddyBench (76.54 vs 76.80 for Opus 5), RaccoonPPTBench v2 (69.60 vs 74.20 for GPT 6 Sol) and BrowseComp (88.55 vs 90.40 for GPT 5.6 Sol). SenseTime notes that DataAnalysis2.0 and RaccoonPPTBench v2 are benchmarks it developed internally. On four memory and agent tests (LongMemEval-S, BEAM 128K, PersonaMem-v2 and π-bench), it does not have the top score on any chart, and it ties GPT 6 Luna on BEAM 128K at 77.71.
TechNode reported on November 3, 2026 that SenseNova 6.8 Flash is available through SenseTime's Raccoon assistant and the SenseNova Token Plan (TechNode).
The release describes access through an API, with a free trial key from the SenseNova Token Plan. It does not mention downloadable weights for SenseNova 6.8 Flash, and the sensenova organization on Hugging Face listed no SenseNova 6.8 Flash weights as of November 9, 2026.
Which SenseNova models are open?
SenseTime publishes open models under the Hugging Face organization "sensenova". As of November 9, 2026, this guide covers three recent releases. The organization also lists earlier checkpoints, including SenseNova-U1, SenseNova-SI and SenseNova-MARS models.
SenseNova-U1.5-8B-MoT (model page) is released under the Apache 2.0 license. SenseTime describes it as a "native unified multimodal" model for image generation and image editing, meaning one model that both understands images and makes them. It is built on an architecture SenseTime calls NEO-unify. The model page lists better text rendering in Chinese and English, infographics, and native 4K generation. A follow-up, SenseNova-U1.5-8B-MoT-TailSteer, targets more stable multi-turn editing.
SenseNova-Loop-9B-Preview (model page) is also Apache 2.0 and was released on October 30, 2026 according to its card. SenseTime calls it "a multimodal language model developed for long-context and agentic workflows". It takes text and images as input and outputs text. The design is "looped": a block of decoder layers runs twice with shared weights. Showcase examples on the card are credited "Created by Raccoon Work".
SenseNova-6.8-Code2Video (model page) is, in SenseTime's words, "a lightweight open model for code-driven video creation". Its card says it builds and renders videos by generating executable code and currently focuses on short videos made entirely in code. We covered it in SenseNova 6.8 Code2Video. Like the other two, it uses the Apache 2.0 license.
These open releases sit alongside, not inside, SenseNova 6.8 Flash. Someone evaluating SenseNova for a project needs to check which side of that line a given model falls on.

What is Raccoon?
Raccoon, Chinese name Xiao Huanxiong (小浣熊), is SenseTime's assistant app brand. It began with an office and data-analysis assistant (Office Raccoon) and a coding assistant called Code Raccoon. On August 25, 2026, SenseTime merged Office Raccoon and Code Raccoon into one product, Raccoon Work, according to a SenseTime release in Korean. That release describes Raccoon Work as a desktop AI agent that runs on the user's PC, with desktop, web and iOS versions.
Raccoon is one of the channels through which SenseTime distributes its models, alongside the API platform. Model credits on Hugging Face also point back to Raccoon for showcase examples.
How can you try SenseNova?
There are four routes as of November 2026.
The API. SenseTime runs a developer platform at platform.sensenova.ai.
The free trial. SenseTime's release describes a free trial that provides an API key through the SenseNova Token Plan, with "up to 1,500 free API requests every five hours" as of November 2026.
The apps. TechNode reported that SenseNova 6.8 Flash is available through the Raccoon assistant, which is the option for readers who want an interface instead of code.
Open downloads. The Apache 2.0 models on Hugging Face can be downloaded and used under the terms of that license, which means running them on your own hardware.
How does SenseNova compare with other Chinese AI labs?
llms.blog has explained several other Chinese model families, including DeepSeek, Qwen, Kimi, GLM, MiniMax and ERNIE, in China's AI labs explained. Qwen comes from Alibaba (What is Qwen) and ERNIE comes from Baidu (What is ERNIE).
Two points distinguish SenseTime from that group. Its company page lists Vision AI next to Generative AI as a core business, and image, video and 3D work appear in SenseNova releases alongside text. Its stated emphasis in the SenseNova 6.8 Flash release is office agents and design output, such as presentations, posters, web pages and 3D scenes.
The second point is the split between closed and open. SenseNova includes API and app models with no published weights, plus Apache 2.0 models aimed at image generation, long-context work and code-driven video. Details on the other labs are in the linked explainers.
Key takeaways
- SenseNova is SenseTime's model series name, unveiled on April 10, 2023. Xu Li said the name implies "constant renewal, daily renewal, and further renewal", and the Chinese name, Ri Ri Xin (日日新), translates literally as "day day new".
- SenseTime was founded in 2014, describes itself as an AI software company, and is listed on The Stock Exchange of Hong Kong. Its site lists 40,400 PFlops of SenseCore compute capacity.
- SenseNova 6.8 Flash, in a release dated November 5, 2026, is what SenseTime calls a "multimodal agent model" for office work and design output. It is served through an API and, per TechNode, the Raccoon assistant, not as an open download. SenseTime's own charts show it ahead on some named tests and behind rivals on others.
- Open Apache 2.0 models on Hugging Face include SenseNova-U1.5-8B-MoT for image generation and editing, SenseNova-Loop-9B-Preview for long-context and agentic workflows, and SenseNova-6.8-Code2Video.
- Access runs through platform.sensenova.ai, the Raccoon Work app, or open model downloads, with a SenseNova Token Plan trial offering "up to 1,500 free API requests every five hours" as of November 2026.
Sources
- SenseTime: About
- SenseTime: SenseNova launch release (April 2023)
- SenseTime: SenseNova 6.8 Flash release (Nov 5, 2026)
- TechNode: SenseNova 6.8 Flash (Nov 3, 2026)
- SenseTime: Office Raccoon opens to Hong Kong and Macau users (Feb 2024, in Chinese)
- SenseTime: Office Raccoon and Code Raccoon merge into Raccoon Work (Aug 25, 2026, in Korean)
- Hugging Face: sensenova organization
- SenseNova-U1.5-8B-MoT model card
- SenseNova-Loop-9B-Preview model card
- SenseNova-6.8-Code2Video model card
