SenseNova-MARS - SenseTime's open-source multimodal autonomous inference model

What is SenseNova-MARS?

SenseNova-MARS is an open-source multimodal autonomous reasoning model from SenseTime, offering two parameter scales: 8B and 32B. As the first Agentic VLM to support deep integration of dynamic visual reasoning and image-text search, the model can autonomously plan task steps like an intelligent agent, flexibly calling upon three major tools: image cropping, text search, and image search, and completing complex multi-hop reasoning without human intervention. In seven benchmark tests, including MMSearch, HR-MMSearch, and FVQA, SenseNova-MARS-32B achieved an average score of 69.74, surpassing Gemini-3-Pro (69.06) and GPT-5.2 (67.64), reaching state-of-the-art (SOTA) level among open-source models, marking a significant leap for AI from "being able to answer" to "being able to execute."

Main functions of SenseNova-MARS

• Multimodal search reasoningThe model integrates images and text for cross-modal information retrieval and supports dynamic invocation of tools to complete complex knowledge reasoning.
• Fine-grained visual analysisIt supports processing 4K/8K ultra-high-definition images and can accurately crop and identify tiny visual details that account for less than 5% of the image.
• Autonomous Agent ExecutionThe model can autonomously plan task steps, seamlessly collaborate with multiple tools to form a closed loop, and solve complex problems without human intervention.

Technical principles of SenseNova-MARS

• Two-stage training architectureThe first stage involves cold-start supervised fine-tuning using approximately 3,000 high-quality multi-round interaction trajectories to enable the model to master the basic tool usage patterns. The second stage uses the BN-GSPO reinforcement learning algorithm to stabilize the training process through in-group and batch normalization, thereby optimizing the collaborative calling and inference decision-making capabilities of multiple tools.
• BN-GSPO reinforcement learning algorithmTo address the training instability caused by differences in trajectory length and reward scale in multi-tool scenarios, the algorithm first performs group normalization on samples in the same group to eliminate internal bias, and then normalizes the entire batch to balance the learning signals of different tasks, thereby achieving stable and efficient multi-tool strategy optimization.
• Motion space and tool designEach round of the model allows users to choose from four actions: text search, image search, image cropping, and termination of response. Image cropping achieves precise local magnification by normalizing bounding box coordinates. All tool calls adhere to strict JSON format specifications to ensure consistent interaction.
• Reward Model MechanismThe model uses GPT-4o as the evaluator, providing sparse rewards from two dimensions: answer accuracy and format compliance. The accuracy reward measures the semantic matching degree between the final answer and the ground truth, while the format reward ensures that each round of output contains a compliant thought process and tool call structure.
• Automated data synthesisBased on a multimodal intelligent agent, a data engine is built. Through three steps—fine-grained visual anchor point localization, multi-hop deep association retrieval, and closed-loop self-consistency verification—it automatically mines cross-web page entity logic and constructs highly complex reasoning links, while filtering illusory data to ensure the quality of training data.

Project address for SenseNova-MARS

• GitHub repositoryhttps://github.com/OpenSenseNova/SenseNova-MARS
• HuggingFace model library:
  • https://huggingface.co/sensenova/SenseNova-MARS-32B
  • https://huggingface.co/sensenova/SenseNova-MARS-8B
• arXiv technical paperhttps://arxiv.org/pdf/2512.24330

Application scenarios of SenseNova-MARS

• Sports competition analysisThe model can identify tiny logos on racing suits and query relevant company and driver background information, and automatically calculate time differences to assist in verifying race data.
• Business intelligence miningIdentify company logos from summit photos to quickly gather product parameters and financing information, aiding in the analysis of the industry's competitive landscape.
• News Fact Check: To trace the background of events and the identities of people through high-resolution news images, and to verify the authenticity of information circulating on social media.
• Academic Research and EducationIt automatically analyzes the data in the figures and tables of papers and retrieves relevant research background, accelerating the literature review and knowledge integration process.
• Geographical travel explorationIt can identify landmarks, road signs, and other details, and retrieve historical and cultural information in real time, providing an immersive smart tour guide experience.