SenseNova-MARS

model paper dataset
SenseTime
Multimodal Agentic Reasoning and Search framework. First Agentic VLM integrating dynamic visual reasoning with text-image search via reinforcement learning. Proposes BN-GSPO algorithm for stable tool-use training. MARS-32B scores 74.3 on MMSearch and 54.4 on HR-MMSearch, surpassing Gemini-3-Pro and GPT-5.2.
Paper (arXiv)GitHubHuggingFace (32B)HuggingFace (8B)HR-MMSearch Benchmark

Outputs 3

SenseNova-MARS Models

model

Variants

Name Parameters Notes
SenseNova-MARS-8B 8B —
SenseNova-MARS-32B 32B —

SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning

paper

arXiv: 2512.24330

HR-MMSearch Benchmark

dataset

First search-oriented benchmark with high-resolution images and knowledge-intensive, search-driven questions.

multimodalreasoningagenticopen-weightevaluation