SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

SenseNova-MARS - SenseTime's Open-Source Multimodal Autonomous Reasoning Model

Author: AI小集 · Original source: AI工具集

What is SenseNova-MARS

SenseNova-MARS is an open-source multimodal autonomous reasoning model from SenseTime, available in two parameter scales: 8B and 32B. As the first Agentic VLM that supports deep integration of dynamic visual reasoning and image-text search, the model can autonomously plan task steps like an agent, flexibly invoking three major tools—image cropping, text search, and image search—and can complete complex multi-hop reasoning without human intervention. Across seven major benchmarks including MMSearch, HR-MMSearch, and FVQA, SenseNova-MARS-32B surpasses with an average score of 69.74Gemini-3-Pro(69.06 points) and GPT-5.2 (67.64 points), reaching SOTA level among open-source models and marking an important leap for AI from "being able to answer" to "being able to execute."

SenseNova-MARS

Main Features of SenseNova-MARS

  • Multimodal Search Reasoning: The model integrates images and text for cross-modal information retrieval and supports dynamic tool invocation to complete complex knowledge reasoning.
  • Fine-Grained Visual Analysis: Supports processing 4K/8K ultra-high-definition images and can precisely crop and recognize tiny visual details accounting for less than 5%.
  • Autonomous Agent Execution: The model can autonomously plan task steps, seamlessly collaborate across multiple tools to form a closed loop, and solve complex problems without human intervention.

Technical Principles of SenseNova-MARS

  • Two-Stage Training Architecture: In the first stage, cold-start supervised fine-tuning is performed using approximately 3,000 high-quality multi-turn interaction trajectories, enabling the model to master basic tool usage patterns; in the second stage, the BN-GSPO reinforcement learning algorithm is used to stabilize the training process through dual normalization within groups and across batches, optimizing multi-tool collaborative invocation and reasoning decision-making capabilities.
  • BN-GSPO Reinforcement Learning Algorithm: To address 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, then normalizes the entire batch to balance the learning signals of different tasks, achieving stable and efficient multi-tool policy optimization.
  • Action Space and Tool Design: In each round, the model can choose among four actions: text search, image search, image cropping, and terminating the answer. Image cropping achieves precise local zoom-in through normalized bounding box coordinates, and all tool calls follow strict JSON format specifications to ensure interaction consistency.
  • Reward Model Mechanism: The model adoptsGPT-4oas the judge, providing sparse rewards from two dimensions: answer accuracy and format compliance. The accuracy reward measures the semantic match between the final answer and the ground truth, while the format reward ensures that each round's output contains a compliant thought process and tool invocation structure.
  • Automated Data Synthesis: Based on a multimodal agent, a data engine is built. Through three stages—fine-grained visual anchor localization, multi-hop deep association retrieval, and closed-loop self-consistency verification—it automatically mines cross-webpage entity logic and constructs highly complex reasoning chains, while filtering hallucinated data to ensure training data quality.

SenseNova-MARS Project Address

  • GitHub Repository：https://github.com/OpenSenseNova/SenseNova-MARS
  • HuggingFace Model Hub：
    • https://huggingface.co/sensenova/SenseNova-MARS-32B
    • https://huggingface.co/sensenova/SenseNova-MARS-8B
  • arXiv Technical Paper：https://arxiv.org/pdf/2512.24330

Application Scenarios of SenseNova-MARS

  • Sports Competition Analysis: The model can identify tiny logos on racing suits and query related company and driver background information, automatically calculating time differences to assist in verifying race data.
  • Business Intelligence Mining: Identify corporate logos from summit photos, quickly gather product specifications and financing information, and assist in analyzing the industry competitive landscape.
  • News Fact-Checking: Trace event background and personal identities from high-resolution news images, and verify the authenticity of information circulating on social media.
  • Academic Research and Education: Automatically analyze data in paper charts and retrieve related research background, accelerating the literature review and knowledge integration process.
  • Geographic Travel Exploration: Recognize details such as landmarks and street signs and retrieve historical and cultural information in real time, providing an immersive intelligent guided tour experience.