SenseNova V6.5

What it is

SenseNova V6.5 is an enhanced version of SenseTime's SenseNova model. SenseTime announced it at "AI+: Large Model Shapes the Future", a forum the company organised at WAIC 2025, and dated the announcement 31 July 2025.^[1] Like other multimodal models, it handles images as well as text; what SenseTime puts at the centre of this release is an interleaved image-text chain of thought.

A chain of thought is a model's habit of writing out intermediate steps before giving an answer, which generally makes multi-step problems easier to solve and easier to check. "Interleaved" means those steps can alternate between pictures and words: the model can work over an image, say something about it in text, return to the image, and carry on. SenseTime said this made V6.5 China's first commercial-grade model with true image-text interleaved reasoning.^[1]

Design and training

Two architectural choices are described. The visual encoder — the part that turns an image into representations the rest of the model can use — was significantly streamlined, and the backbone was made deeper and narrower, with visual and language representations aligned early.^[1] Aligning early means the two modalities are pushed into a shared representation near the start of the network rather than being reconciled only at the end, which tends to make cross-modal reasoning cheaper to learn.

Training ran in stages: seed data built on an understanding of cognitive processes, then supervised fine-tuning, then multiple rounds of reinforcement learning.^[1] Supervised fine-tuning trains a model on examples of the behaviour wanted; reinforcement learning then rewards outputs that score well against some objective, a common way to sharpen reasoning after the initial training run.

Reported performance

SenseTime reported pre-training throughput more than 20% higher, reinforcement-learning efficiency up 40%, and inference throughput more than 35% higher.^[1] It put the cost-performance ratio at three times that of V6.0.^[1] On capability, the company claimed that V6.5's text and multimodal reasoning surpassed Gemini 2.5 Pro and Claude 4 Sonnet, and that its multimodal interaction surpassed Gemini 2.5 Flash and GPT-4o.^[1] These are SenseTime's own comparisons.

In products

V6.5 powers an agent upgrade of Raccoon, SenseTime's assistant, which the company said has more than 10 million users.^[1] SenseTime said the Raccoon product suite has been widely adopted across a range of industries, and that Raccoon significantly outperforms models such as OpenAI's o3 in data analysis.^[1] An agent upgrade means the assistant is built to carry out multi-step tasks rather than only answer single questions.

Where it sits

V6.5 builds on V6.0, and the release is framed largely around efficiency and cost: higher throughput, less cost per unit of work, and reasoning that mixes images and text in a single stream. SenseTime described the step as a move from AI as a "tool" to a "partner".^[1] The comparisons are the company's own; how the model stands will depend on independent evaluation and on how it performs inside the Raccoon products.