# SenseNova > Source: https://aiwiki.ai/wiki/sensenova > Updated: 2026-11-01 > Fact-checked: 2026-10-29 > Categories: AI Models, Chinese AI, Large Language Models, Multimodal AI > License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)" > Cite as: AI Wiki. "SenseNova." aiwiki.ai, 1 Nov 2026. https://aiwiki.ai/wiki/sensenova > From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution. **SenseNova** (Chinese: 日日新) is the family of [foundation models](https://aiwiki.ai/wiki/foundation_models) developed by the Hong Kong-listed AI company [SenseTime](https://aiwiki.ai/wiki/sensetime). SenseTime introduced the name on 10 April 2023, when it presented SenseNova as a "foundation model set" covering natural language processing, content generation, automated data annotation and custom model training, with the SenseChat [large language model](https://aiwiki.ai/wiki/large_language_model) as its conversational front end.[1][2] Chairman and CEO [Xu Li](https://aiwiki.ai/wiki/xu_li) said the name implied "constant renewal, daily renewal, and further renewal".[1] The brand now spans SenseTime's proprietary flagship generations (4.0 through V6.5, and the 6.x agent models), the products built on them such as the Office Raccoon assistant, and a set of open-weight models (Chinese embedding models from 2023, and a much larger group of research models from late 2025 onward) published under the "sensenova" organization on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) and the OpenSenseNova organization on [GitHub](https://aiwiki.ai/wiki/github).[3][4]
SenseTime logo
SenseTime logo. SenseTime, via Wikimedia Commons, CC BY-SA 4.0.
Since 2025 the line has moved away from adapter-based [multimodal models](https://aiwiki.ai/wiki/multimodal_model) toward what SenseTime calls native multimodal architectures: NEO (December 2025) and NEO-unify (March 2026), the basis of the open SenseNova U1 and U1.5 models, which handle image understanding and image generation in one network without a separate vision encoder or VAE.[5][6][7] ## History ### 2023: launch of SenseNova and SenseChat SenseTime unveiled SenseNova at a Tech Day event; the company later dated the launch to 10 April 2023, and the PR Newswire release carries an 11 April dateline.[1][2] The release described SenseChat as an LLM "with hundreds of billions of parameters" trained with a focus on Chinese text, and demonstrated a programming assistant, a health consultation assistant and a PDF reading assistant built on it, along with text-to-image creation and 2D/3D digital human generation.[1] SenseTime presented the models as running on SenseCore, its AI infrastructure, which it said comprised 27,000 GPUs delivering 5,000 petaflops.[1] ### 2024: SenseNova 4.0, 5.0, 5.5 and 5o SenseNova 4.0 was announced on 6 February 2024.[8] Its general LLM, SenseChat V4, supported 4k, 32k and 128k token context windows. SenseTime said its overall performance "closely matches that of GPT-4" on the OpenCompass evaluation platform and reported a HumanEval one-pass rate of 75.6% against 74.4% for [GPT-4](https://aiwiki.ai/wiki/gpt_4).[2] The same release added SenseChat-Vision V4 (a 30-billion-parameter multimodal model), SenseChat-Medical V4, the code-interpreter model SenseChat-DataAnalysis V4, the 10-billion-parameter text-to-image model SenseMirage V4, a SenseChat Function Call & Assistants API, and the first version of Office Raccoon, a no-code data-analysis tool. SenseTime said SenseNova had attracted more than 3,000 enterprise users since its launch.[2] SenseNova 5.0 followed on 23 April 2024 at a Tech Day event in Shanghai. SenseTime said it had been trained on more than 10TB of tokens, including a large amount of synthetic data, used a [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts) design, and reached an effective [context window](https://aiwiki.ai/wiki/context_window) of about 200,000 tokens at inference. The company claimed first place on the MMBench multimodal benchmark and launched edge-side models at the same event. Partners named in the release included Kingsoft Office (WPS 365), Haitong Securities, and Xiaomi, whose SU7 car uses SenseTime large-model technology in its smart cabin.[9] At the World Artificial Intelligence Conference in Shanghai in July 2024, SenseTime presented SenseNova 5.5 and SenseNova 5o.[10] The company claimed 5.5 was 30% better overall than 5.0, with mathematical reasoning up 31.5%, English up 53.8% and prompt following up 26.8%. It described 5o as China's first real-time multimodal model, taking audio, text, image and video input in conversation, and compared it with [GPT-4o](https://aiwiki.ai/wiki/gpt_4o). An edge model, SenseChat Lite-5.5, and the Vimi avatar video generator were shown at the same event.[10] ### 2025: V6, V6.5 and the NEO architecture SenseNova V6 was launched at Tech Day events held in several cities, including Shanghai and Shenzhen, in April 2025 (the PR Newswire release is dated 12 April).[11] SenseTime described it as a native MoE multimodal general foundation model with more than 600 billion parameters, trained on over 200B of high-quality multimodal long [chain-of-thought](https://aiwiki.ai/wiki/chain_of_thought) data, with reasoning chains up to 64K tokens. The company said it was the first model in China to understand and reason over videos about 10 minutes long, benchmarked its reasoning against GPT-4.5 and Gemini 2.0 Pro, and claimed the lowest training and inference costs in the industry. The release also covered V6 Reasoner and SenseNova V6 Omni, a lightweight real-time interactive model that SenseTime called China's first commercialized full-modality real-time interactive model.[11] SenseNova V6.5 was launched at a SenseTime-organized large-model forum during WAIC 2025 in July 2025.[12][5] Its main addition was interleaved image-text "multimodal thought chains", which SenseTime described as the first commercial-grade image-text interleaved reasoning in China. The company said a streamlined visual encoder and a deeper, narrower backbone raised pretraining throughput by over 20%, reinforcement-learning efficiency by 40% and inference throughput by over 35%, for a threefold cost-performance gain over V6.0. It claimed text and multimodal reasoning ahead of [Gemini 2.5 Pro](https://aiwiki.ai/wiki/gemini_2_5_pro) and Claude 4 Sonnet.[12] On 2 December 2025 SenseTime released and open-sourced NEO, a native vision-language architecture co-developed with S-Lab of Nanyang Technological University, which it called the cornerstone of the next-generation SenseNova multimodal models.[5] The accompanying paper, "From Pixels to Words: Towards Native Vision-Language Primitives at Scale" (arXiv 2510.14979, accepted at ICLR 2026), describes NEO as a family of native vision-language models that learn visual perception from scratch inside "a dense and monolithic model", trained on 390 million image-text examples, instead of attaching a separate vision module to a language model.[13] SenseTime said that data volume was about one tenth of what modular models of equivalent performance need, and that NEO matched models such as Qwen2-VL and InternVL3 on several visual understanding tasks.[5] ### 2026: NEO-unify, U1, U1.5 and the 6.x agent models In March 2026 SenseTime and Nanyang Technological University described NEO-unify, an architecture that removes both the vision encoder and the VAE and uses a native Mixture-of-Transformers (MoT) backbone so that understanding and generation run in one model.[6] SenseNova U1, the first models built on it, had weights posted on Hugging Face on 27 April 2026, and SenseTime announced the release on 30 April.[3][7] SenseTime called this open release the U1 Lite series; it said the U1 Lite models' general image generation was comparable to Qwen-Image 2.0 Pro and Seedream 4.5 and that it could produce "continuous image-text" (interleaved) output in a single call.[7] A distilled U1 Fast variant dedicated to infographic generation runs on the free SenseNova-Studio playground.[3] On 8 May 2026 SenseTime released SenseNova 6.7 Flash-Lite, a lightweight multimodal agent model for office work such as data analysis, slide generation, research reports and infographics. It is offered through an API rather than as open weights. SenseTime said it cut token use by about 60% compared with text-only agents in tasks such as information search. The launch came with a limited-time free SenseNova Token Plan and the open-source SenseNova-Skills office skill set for agent frameworks such as [OpenClaw](https://aiwiki.ai/wiki/openclaw) and Hermes Agent.[14][15] On 18 July 2026 SenseTime announced SenseNova U1 Pro, an image-creation model that uses an internal interleaved image-text chain of thought and outputs up to native 8K resolution.[16] A production version became available through the Raccoon app and the SenseNova API in September 2026.[17] In August 2026 the company published SenseNova 6.8 Flash Lite Preview through the Token Plan, an early preview aimed at long-horizon "delegated" tasks in which a lead agent dispatches more than ten specialist sub-agents. As of late October 2026 its README still says SenseNova 6.8 Flash Lite and a larger SenseNova 6.8 Flash are "coming soon",[18] although SenseTime's website FAQ, accessed 29 October 2026, names SenseNova 6.8 Flash Lite as a featured model and one of the models supported by the Token Plan.[39] SenseTime's first-half 2026 results listed the 2026 model releases as the NEO-unify architecture, SenseNova U1, U1.5 and U1.5-Lite, SenseNova Vision, 6.8-Flash-Lite, U1-Pro, the spatial intelligence model SenseNova SI-8B and the Kairos 3.1 world model co-developed with ACE Robotics.[19] SenseNova U1.5 Lite went on the Token Plan API on 25 August 2026 with support for output up to 4096 x 4096.[20] Open SenseNova-U1.5-8B-MoT weights, preceded by a preview checkpoint, were published on Hugging Face, and the technical report (arXiv 2609.11929) was submitted on 10 September 2026.[21][22] In October 2026 the sensenova Hugging Face organization added an embodied-agent model, SenseNova-RoboRSI-27B-Preview, and SenseNova 6.8 Code2Video, a code-driven video creation model.[23][24] ## Flagship and API releases | Release | Date | Description (as described by SenseTime) | Open weights | |---|---|---|---| | SenseNova / SenseChat | 10 April 2023 | Foundation model set; SenseChat LLM with "hundreds of billions of parameters"[1][2] | No | | SenseNova 4.0 | 6 February 2024 | SenseChat V4 (4k/32k/128k context), Vision, Medical, DataAnalysis, SenseMirage V4[2][8] | No | | SenseNova 5.0 | 23 April 2024 | MoE LLM, 10TB+ tokens, ~200K effective context[9] | No | | SenseNova 5.5 and 5o | July 2024 | 30% overall gain claimed over 5.0; 5o real-time multimodal model[10] | No | | SenseNova V6 (V6 Reasoner, V6 Omni) | April 2025 | Native MoE multimodal model, 600B+ parameters, 64K chain of thought[11] | No | | SenseNova V6.5 | July 2025 | Interleaved image-text chain of thought; 3x cost-performance over V6.0[12] | No | | SenseNova U1 (Lite series) | April 2026 | NEO-unify unified understanding and generation; 8B-MoT and A3B-MoT[3][7] | Yes (Apache 2.0) | | SenseNova 6.7 Flash-Lite | 8 May 2026 | Lightweight multimodal agent model for office workflows[14][15] | No (API) | | SenseNova U1 Pro | 18 July 2026 (production September 2026) | Image creation model, up to 8K output[16][17] | No (API, Raccoon app) | | SenseNova 6.8 Flash Lite Preview | August 2026 | Long-horizon, multi-agent "delegated intelligence" preview[18] | No (API) | | SenseNova U1.5 / U1.5 Lite | August-September 2026 | Unified model with native 4K generation and editing[20][21] | 8B-MoT yes (Apache 2.0); Lite via API | | SenseNova 6.8 Code2Video | October 2026 | Code-driven video creation model fine-tuned from Qwen3.8-27B[24] | Yes (Apache 2.0) | ## Open-weight models SenseTime's open releases sit on the Hugging Face organization "sensenova" and the GitHub organization OpenSenseNova, and are mostly research models from SenseTime Research, often with co-authors at Nanyang Technological University or Tsinghua University.[4][25][26] The table lists the main lines; dates are the earliest dated record found (arXiv submission, model-card news entry, or Hugging Face repository creation), which can precede general availability by a few days. | Model line | Earliest record | What it is | Sizes | License | Paper | |---|---|---|---|---|---| | piccolo-base-zh, piccolo-large-zh | September 2023 | Chinese text [embedding](https://aiwiki.ai/wiki/embedding) models trained on 400 million weakly supervised and 20 million labeled text pairs[27] | 768- and 1024-dimension | MIT | None | | piccolo-large-zh-v2 | April 2024 | Chinese embedding model, 1792-dimension, multi-task hybrid loss[28][29] | 1.21 GB | Not stated | arXiv 2405.06932 | | InteractiveOmni | October 2025 | Omni-modal model for audio-visual multi-turn dialogue with speech output[30] | 4B, 8B | MIT | arXiv 2510.13747 | | SenseNova-SI | November 2025 | [Spatial intelligence](https://aiwiki.ai/wiki/spatial_intelligence) fine-tunes of InternVL3, Qwen3-VL, Qwen2.5-VL and BAGEL[25][31] | 2B-8B | Apache 2.0 | arXiv 2511.13719 | | SenseNova-MARS | December 2025 | Agentic multimodal reasoning and search trained with reinforcement learning[26] | 8B, 32B | MIT | arXiv 2512.24330 | | ConsistCompose | February 2026 (model) | Layout-controlled multi-instance image composition on BAGEL-7B-MoT[32] | 7B MoT | Apache 2.0 | arXiv 2511.18333 | | SenseNova-U1 | April-May 2026 | Native unified understanding and generation on NEO-unify[3][33] | 8B-MoT (dense), A3B-MoT (MoE) | Apache 2.0 | arXiv 2605.12500 | | SenseNova-Vision | June-July 2026 | Computer vision tasks recast as text, image or mixed generation[34] | 7B MoT | CC BY-NC 4.0 | arXiv 2607.06560 | | SenseNova-U1.5 | July-September 2026 | Successor to U1 with native resolutions up to 4K[21][22] | 8B-MoT | Apache 2.0 | arXiv 2609.11929 | | Looped-DiT | September 2026 | Text-to-image [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) that loops shared blocks within each denoising step[35] | B/16, B/32 | MIT | None yet | | SenseNova-RoboRSI-27B-Preview | October 2026 | Vision-language model for embodied agents, fine-tuned from Qwen3.8-27B[23] | 27B | Apache 2.0 | Report "coming soon" | | SenseNova 6.8 Code2Video | October 2026 | Code-driven video creation, fine-tuned from Qwen3.8-27B[24] | 27B | Apache 2.0 | None | ### SenseNova-U1 and U1.5 The U1 model card describes "8B-MoT" as about 8 billion understanding parameters plus about 8 billion generation parameters. SenseNova-U1-8B-MoT uses a dense backbone and SenseNova-U1-A3B-MoT a mixture-of-experts backbone; the technical report says they are built on dense 8B and 30B-A3B understanding baselines respectively.[3][33] Released checkpoints include SFT and final versions (the final models add a round of text-to-image reinforcement learning), an 8-step distilled preview, LoRAs and infographic-tuned variants. The card lists limits: a 32K-token context for visual understanding, unstable fine human details, occasional text-rendering errors, and interleaved generation that is still experimental.[3] The authors also show preliminary vision-language-action and [world model](https://aiwiki.ai/wiki/world_model) uses.[33] U1.5 keeps the 8B-MoT, encoder-free and VAE-free design. Its report describes spatially coherent patch reconstruction, training at native resolutions up to 4K, separate expert models for aesthetics, bilingual text rendering, infographics and image editing, and merging of those experts through multi-expert on-policy distillation. The authors say they will release training code for supervised fine-tuning, reinforcement learning and on-policy distillation.[22] ### SenseNova-SI SenseNova-SI is a series of models fine-tuned for spatial reasoning on SenseNova-SI-8M, a curated set of eight million samples organized by a taxonomy of spatial capabilities. The base models include InternVL3, Qwen3-VL, Qwen2.5-VL and BAGEL.[31][25] According to the model zoo on the card, later InternVL3-8B releases scaled the spatial training data to 10 million (1.2), 14 million (1.3) and 29 million (1.4) samples, and version 1.5 added 1.5 million samples on top of 1.4.[25] The GitHub repository lists the work as a CVPR 2026 paper.[4] ### SenseNova-MARS, InteractiveOmni and SenseNova-Vision SenseNova-MARS (Multimodal Agentic Reasoning and Search) trains a [vision-language model](https://aiwiki.ai/wiki/vision_language_model) to interleave image search, text search and image cropping with its reasoning. It uses a reinforcement-learning algorithm the authors call Batch-Normalized Group Sequence Policy Optimization (BN-GSPO), and the paper introduces the HR-MMSearch benchmark of high-resolution, search-driven questions. The 8B model is fine-tuned from Qwen3-VL-8B-Instruct.[26][36] InteractiveOmni combines a vision encoder, audio encoder, LLM and speech decoder in a single model; its authors report that the 4B version keeps 97% of the 8B version's performance at half the size.[30] SenseNova-Vision expresses detection, OCR, keypoints, segmentation, depth, surface normals, point maps and camera pose estimation as outputs of a unified generation model, with no task-specific heads.[34] ### SenseNova-RoboRSI and Code2Video SenseNova-RoboRSI-27B-Preview is an open-weight 27B model for [embodied AI](https://aiwiki.ai/wiki/embodied_ai) agents trained through what SenseTime calls joint model and harness recursive self-improvement on validated interaction experience; the card covers benchmark reproduction on LIBERO-Pro and RoboCasa365, with a technical report still to come.[23] SenseNova 6.8 Code2Video builds and renders short videos by writing executable code and is designed to run for more than 10 hours and 2,000 turns inside an agent harness such as [Claude Code](https://aiwiki.ai/wiki/claude_code), with a 512K-token context. The release includes weights, a production skill and example trajectories. It is covered in more detail at [SenseNova 6.8 Code2Video](https://aiwiki.ai/wiki/sensenova_6_8_code2video).[24] ## Architecture: NEO and NEO-unify Most multimodal LLMs, including earlier SenseNova generations, connect a pretrained vision encoder to a language model through a projector or adapter. SenseTime's NEO work drops that modular design. According to SenseTime, NEO uses a patch embedding layer that maps pixels directly to tokens, a native 3D rotary position embedding (Native-RoPE) with separate frequency allocation for visual and text dimensions, and attention that is autoregressive for text tokens and bidirectional for visual tokens within one framework.[5][13] NEO-unify extends this to generation. It removes the variational autoencoder used by most image generators as well as the vision encoder, uses a near-lossless visual interface for both input and output, and splits work between understanding and generation branches inside a Mixture-of-Transformers backbone. In a SenseTime experiment, a 2B NEO-unify model reached 31.56 PSNR and 0.85 SSIM on MS COCO 2017 image reconstruction after 90,000 pretraining steps, against 32.65 and 0.91 for the Flux VAE.[6] SenseNova U1 and U1.5 are the production models built on this design.[7][22] ## Products and access - **SenseChat**: the conversational LLM introduced with SenseNova in 2023 and versioned with each generation (for example SenseChat V4 and SenseChat Lite-5.5).[1][2][10] - **Office Raccoon**: launched in February 2024 as a natural-language data-analysis tool built on SenseNova.[2] With V6.5 it was upgraded into an office agent with education and finance editions, and SenseTime said in July 2025 that the Raccoon product suite had more than 10 million users.[12] SenseTime's other figures count users differently: its first-half 2025 results said the Raccoon series had surpassed 3 million users, and the December 2025 Raccoon 3.0 announcement cited more than 3 million registered users serving more than 15 million individuals.[40][41] In its first-half 2026 results, SenseTime named Lenovo, Ping An Technology, China's three major telecom operators, JD.com and Kylin Software as Raccoon customers.[19] - **API platform and Token Plan**: SenseNova models are offered through the platform.sensenova.cn developer platform. The Token Plan, launched with 6.7 Flash-Lite in May 2026, initially gave developers 1,500 calls every five hours in the first month.[14] SenseTime said in August 2026 that average daily Token Plan usage had passed 500 billion tokens.[20] - **SenseNova-Skills and SenseNova-Studio**: an open-source skill library that packages SenseNova models for agent frameworks, and a free browser playground for the U1 models.[3][15] ## Benchmark claims The OpenCompass and SuperCLUE rows are third-party leaderboard results as relayed by SenseTime; all other figures are self-reported by SenseTime or the model's authors and have not been independently replicated. | Model | Benchmark | Reported result | Source | |---|---|---|---| | SenseChat V4 | HumanEval (one pass) | 75.6% (GPT-4: 74.4%) | SenseTime, February 2024[2] | | SenseChat-Vision V4 | MME | 2199.5 (GPT-4: 1926.57) | SenseTime, February 2024[2] | | SenseNova V6.5 Pro | OpenCompass multimodal academic leaderboard | 82.2 overall, ranked first | SenseTime, September 2025[37] | | SenseNova V6.5 Pro | SuperCLUE Chinese multimodal VLM evaluation (December 2025 report) | 75.35 overall, first among Chinese models | SenseTime, January 2026[38] | | SenseNova-SI | VSI-Bench / MMSI / MindCube | 68.8% / 43.3% / 85.7% | Paper authors[31] | | SenseNova-MARS-32B | MMSearch / HR-MMSearch | 74.3 / 54.4 | Model card[36] | | SenseNova-RoboRSI-27B-Preview | LIBERO-Pro (VLA-assisted / direct VLM) | 92.5% / 95.0% | Model card[23] | | SenseNova-RoboRSI-27B-Preview | RoboCasa365 success rate | 46.5% (Qwen3.8-27B base: 26.0%) | Model card[23] | | piccolo-large-zh-v2 | C-MTEB average (35 tasks) | 70.95 | Model card[28] | | Looped-DiT B/16 | GenEval / DPG-Bench | 87.4 / 87.0 | Model card[35] | On OpenCompass, SenseTime said V6.5 Pro outscored Gemini 2.5 Pro and GPT-5; on SuperCLUE it said V6.5 Pro was ahead of Doubao-seed-1.6-vision, ERNIE-5.0 and Qwen3 but behind leading international models.[37][38] The MARS card says the 32B model surpassed Gemini-3-Pro and GPT-5.2 on its two search benchmarks, and the RoboRSI card ranks the model second to [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) on LIBERO-Pro among the models it compared.[36][23] ## Business context SenseTime reports SenseNova-driven sales within its generative AI business. For the first half of 2026 it reported generative AI revenue of RMB2.33 billion, nearly 80% of total revenue of RMB2.91 billion, and its first IFRS net profit since listing (RMB620 million), which it attributed in part to fair-value gains on ecosystem investments.[19] The same announcement said SenseNova U1 had been adapted on day zero to more than ten Chinese domestic chip platforms and that, less than four months after U1 was open-sourced, three related GitHub projects had collectively passed 11,000 stars.[19] Company-level financials, sanctions history and corporate details are covered in the [SenseTime](https://aiwiki.ai/wiki/sensetime) article. ## See also - [SenseTime](https://aiwiki.ai/wiki/sensetime) - [Xu Li](https://aiwiki.ai/wiki/xu_li) - [SenseNova 6.8 Code2Video](https://aiwiki.ai/wiki/sensenova_6_8_code2video) - [Artificial intelligence in China](https://aiwiki.ai/wiki/china_ai) - [Multimodal model](https://aiwiki.ai/wiki/multimodal_model) - [Spatial intelligence](https://aiwiki.ai/wiki/spatial_intelligence) ## References 1. "SenseTime Launches 'SenseNova' Foundation Model Sets and AI Computing Systems, Advancing AGI Development." PR Newswire (via AAP), April 11, 2023. https://www.aap.com.au/aapreleases/cision20230410ae65464/ 2. "SenseTime Unveils SenseNova 4.0, Bringing Novel AI Experience." SenseTime, February 8, 2024. https://www.sensetime.com/en/news/51167493/ 3. "SenseNova-U1-8B-MoT" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/SenseNova-U1-8B-MoT 4. OpenSenseNova organization repositories. GitHub, accessed October 29, 2026. https://github.com/OpenSenseNova 5. "Evolving From 'Data Fusion' to 'Native Architecture', SenseTime Releases NEO Architecture Redefining the Efficiency Boundaries of Multimodal Models." SenseTime, December 2, 2025. https://www.sensetime.com/en/news/51170250/ 6. "NEO-unify: A Native End-to-End Architecture for Unified Multimodal Understanding and Generation." SenseTime, March 7, 2026. https://www.sensetime.com/en/news/51170548/ 7. "SenseTime Fully Open-Sources SenseNova U1: A Unified Model for Understanding and Generation." SenseTime, April 30, 2026. https://www.sensetime.com/en/news/51170629/ 8. "SenseTime Unveils SenseNova 4.0, Bringing Novel AI Experience." PR Newswire (via AAP), February 6, 2024. https://www.aap.com.au/aapreleases/cision20240206ae29968/ 9. "SenseTime launches SenseNova 5.0 with comprehensive updates and the industry-leading 'Cloud-to-Edge' full-stack large model product matrix." SenseTime, April 25, 2024. https://www.sensetime.com/en/news/51167731/ 10. "SenseTime unveils SenseNova 5o, China's first real-time multimodal AI model to rival GPT-4o." The Decoder, July 8, 2024. https://the-decoder.com/sensetime-unveils-sensenova-5o-chinas-first-real-time-multimodal-ai-model-to-rival-gpt-4o/ 11. "SenseTime's SenseNova V6: China's Most Advanced Multimodal Model with the Lowest Cost in the Industry." PR Newswire, April 12, 2025. https://www.prnewswire.com/apac/news-releases/sensetimes-sensenova-v6-chinas-most-advanced-multimodal-model-with-the-lowest-cost-in-the-industry-302426998.html 12. "SenseTime Launches the Enhanced SenseNova V6.5: Marking the Leap from AI as a 'Tool' to a 'Partner'." SenseTime, July 31, 2025. https://www.sensetime.com/en/news/51169861/ 13. Diao, H. et al. "From Pixels to Words -- Towards Native Vision-Language Primitives at Scale." arXiv:2510.14979, October 2025. https://arxiv.org/abs/2510.14979 14. "China's SenseTime Releases SenseNova 6.7 Flash-Lite Lightweight Multimodal Model, Reducing Token Consumption by 60%." Wedoany, May 8, 2026. https://en.wedoany.com/shortnews/162333.html 15. "SenseNova 6.7 Flash-Lite" README. OpenSenseNova, GitHub, accessed October 29, 2026. https://github.com/OpenSenseNova/SenseNova6.7 16. "商汤科技发布日日新 U1 Pro 图片创作模型." IT之家 (IThome), July 18, 2026. https://www.ithome.com/0/978/562.htm 17. "SenseTime releases SenseNova U1 Pro image model with up to 8K output." TechNode, September 22, 2026. https://technode.com/2026/09/22/sensetime-releases-sensenova-u1-pro-image-model-with-up-to-8k-output/ 18. "SenseNova 6.8 Flash Lite Preview" README. OpenSenseNova, GitHub, August 2026. https://github.com/OpenSenseNova/SenseNova6.8 19. "SenseTime Records First-Ever Profit in First Half of 2026; 'Models + Token Factory + Agent Harness' Framework Unlocks High-Value Commercialization." SenseTime, August 26, 2026. https://www.sensetime.com/en/news/2026H1-results 20. "Free to try, easy to integrate! SenseNova U1.5 Lite is now available through the Token Plan." SenseTime, August 25, 2026. https://www.sensetime.com/en/news/sensenova-u1-5-lite-token-plan-20260911-1741 21. "SenseNova-U1.5-8B-MoT" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT 22. Diao, H. et al. "SenseNova-U1.5: Towards Native Unified Visual Intelligence." arXiv:2609.11929, September 2026. https://arxiv.org/abs/2609.11929 23. "SenseNova-RoboRSI-27B-Preview" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/sensenova-roborsi-27b-preview 24. "SenseNova 6.8 Code2Video" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/SenseNova-6.8-Code2Video 25. "SenseNova-SI-1.5-InternVL3-8B" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/SenseNova-SI-1.5-InternVL3-8B 26. Chng, Y. X. et al. "SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning." arXiv:2512.24330, December 2025. https://arxiv.org/abs/2512.24330 27. "piccolo-large-zh" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/piccolo-large-zh 28. "piccolo-large-zh-v2" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/piccolo-large-zh-v2 29. Huang, J. et al. "Piccolo2: General Text Embedding with Multi-task Hybrid Loss Training." arXiv:2405.06932, May 2024. https://arxiv.org/abs/2405.06932 30. Tong, W. et al. "InteractiveOmni: A Unified Omni-modal Model for Audio-Visual Multi-turn Dialogue." arXiv:2510.13747, October 2025. https://arxiv.org/abs/2510.13747 31. Cai, Z. et al. "Scaling Spatial Intelligence with Multimodal Foundation Models." arXiv:2511.13719, November 2025. https://arxiv.org/abs/2511.13719 32. Shi, X. et al. "ConsistCompose: Unified Multimodal Layout Control for Image Composition." arXiv:2511.18333, November 2025. https://arxiv.org/abs/2511.18333 33. Diao, H. et al. "SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture." arXiv:2605.12500, May 2026. https://arxiv.org/abs/2605.12500 34. Han, X. et al. "Vision as Unified Multimodal Generation." arXiv:2607.06560, July 2026. https://arxiv.org/abs/2607.06560 35. "Looped-DiT B/16" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/Looped-DiT-B16 36. "SenseNova-MARS-32B" model card. SenseTime (sensenova), Hugging Face, accessed October 29, 2026. https://huggingface.co/sensenova/SenseNova-MARS-32B 37. "SenseTime's SenseNova Claims Top Spot in OpenCompass Global Multi-Modal Model Academic Leaderboard." SenseTime, September 12, 2025. https://www.sensetime.com/en/news/51169948/ 38. "SenseTime SenseNova V6.5 Multimodal Large Model Ranked No.1 in China in 2025." SenseTime, January 13, 2026. https://www.sensetime.com/en/news/51170359/ 39. "Frequently Asked Questions." SenseTime, accessed October 29, 2026. https://www.sensetime.com/en/faq 40. "SenseTime Announces 2025 Interim Results: Revenue Exceeded Market Expectations with Period-over-Period Growth of 36% in 1H 2025, Loss Significantly Narrowed." PR Newswire, August 28, 2025. https://www.prnewswire.com/apac/news-releases/sensetime-announces-2025-interim-results-revenue-exceeded-market-expectations-with-period-over-period-growth-of-36-in-1h-2025-loss-significantly-narrowed-302541138.html 41. "DAY2 | SenseTime Raccoon 3.0 is Here: AI-Powered Office Agent Generates High-Quality PPTs in One Click." SenseTime, December 17, 2025. https://www.sensetime.com/en/news/51170347