SenseTime Claims AI Benchmark Lead Over Google, OpenAI with New Open-Source Model

Shanghai's SenseTime Open-Sources Top-Tier AI Model, Claiming Benchmark Superiority Over Google and OpenAI

In a significant move within the competitive artificial intelligence landscape, Chinese AI giant SenseTime has released its latest multimodal model, SenseNova-MARS, under a full open-source license. The company asserts that the model outperforms leading proprietary systems from Google and OpenAI in key visual reasoning and search benchmarks, potentially shifting dynamics in both the research community and commercial AI application development.

The model, available in two parameter sizes (8 billion and 32 billion), was unveiled today alongside complete access to its code, training data, and model weights on platforms like GitHub and Hugging Face. According to SenseTime's technical documentation, SenseNova-MARS achieved a composite score of 69.74 on a core set of multimodal search and reasoning benchmarks. This score, the company claims, edges out Google's Gemini 3.0 Pro (69.06) and OpenAI's GPT-5.2 (67.64).

This release positions SenseNova-MARS as one of the most capable open-source Vision-Language Models (VLMs) to date, specifically architected for complex, multi-step tasks. It enters a field where open-source models like those from Meta have driven widespread innovation but have often trailed behind the absolute performance frontier guarded by closed, proprietary models from leading U.S. firms.

Performance Claims and Benchmark Leadership

SenseTime's announcement is centered on performance data drawn from several established and challenging benchmarks for multimodal AI. The stated average score of 69.74 is derived from tests including MMSearch, FVQA, InfoSeek, and LiveVQA. The company highlighted particularly strong results in two areas.

On the MMSearch benchmark, a core evaluation for image-text search, SenseNova-MARS is reported to have scored 74.27, significantly higher than the cited GPT-5.2 score of 66.08. Perhaps more notably, the model reportedly leads on the demanding HR-MMSearch benchmark with a score of 54.43. SenseTime describes this test as an "Olympics for AI," designed to push models to their limits.

The HR-MMSearch evaluation employs 305 newly created, high-resolution 4K images from 2025 to prevent models from relying on pre-existing data. Questions focus on minute details constituting less than 5% of an image—such as small logos, fine text, or tiny objects—necessitating the use of image-cropping tools for analysis. The test covers eight domains, including sports, finance, and academic research, with 60% of questions requiring the use of at least three different tools to solve.

"The results, if independently verified, suggest a model with exceptional fine-grained visual understanding and planning capability," said an AI researcher at a European university, who requested anonymity as they had not yet tested the model. "Outperforming Google and OpenAI on such a meticulous benchmark would be a substantial technical achievement for any team, open-source or not."

Technical Core: An "Agentic" Model for Dynamic Reasoning

The key technical differentiator for SenseNova-MARS, according to its developers, is its design as an "agentic" visual language model. Unlike standard VLMs that primarily describe or answer questions about images, an agentic model is designed to autonomously plan sequences of actions and invoke external tools to accomplish a goal.

SenseNova-MARS integrates dynamic visual reasoning with deep image-text search capabilities. In practice, this means the model can decide to first crop an image to zoom in on a detail, then run an image search to identify an object, followed by a text search to retrieve related factual data, all within a single, self-devised chain of reasoning. This allows it to tackle complex, multi-hop queries without human intervention.

SenseTime provided illustrative examples: identifying a tiny logo on a race car driver's suit, searching for the founding year of the associated company, matching the driver's birth date, and calculating the age difference; or analyzing a photo from an industry summit to identify corporate logos and swiftly gather information on products and market parameters.

A Two-Phase Training Methodology

The development of such a capable agentic model addressed the significant challenge of scarce training data for complex, cross-modal reasoning tasks. SenseTime's research team detailed a two-phase training regimen.

The first phase focused on building a foundation. To overcome the data scarcity, the team created an automated data synthesis engine. This engine uses "fine-grained visual anchors" and a "multi-hop deep association retrieval" mechanism to dynamically mine and link entities across web pages, automatically constructing intricate reasoning chains. A self-consistency check was implemented to filter out "hallucinated" or illogical data, resulting in a high-quality dataset of complex search-and-reasoning questions with annotated tool-use steps.

The second phase employed reinforcement learning to hone the model's practical skills. The AI agent learned by receiving rewards for correct decisions (e.g., choosing the right tool, formulating a logical step) and adjusting its strategy for errors. To ensure stable learning across tasks of varying difficulty, the team implemented a stabilization algorithm called BN-GSPO (Bootstrapped Nash Gradient-based Skill-Policy Optimization). This algorithm is described as normalizing learning signals to manage the optimization fluctuations caused by the diverse outcomes of dynamic tool calls, solving convergence challenges in training.

"This approach aims to move beyond mere tool usage to instill what we might call 'tool-use intuition,'" explained a SenseTime research lead in the accompanying technical paper. "The model develops a sense for which tools to deploy in which context and how to synthesize their results."

Implications for Commercial and Research Applications

The open-source release of a model with these claimed capabilities could have broad implications. For developers and enterprises, it provides a powerful, royalty-free base for building advanced AI applications in fields like media analysis, competitive intelligence, academic research, and content moderation, where understanding visual details in context is crucial.

For instance, in financial analysis, the model could automatically extract and cross-reference data from complex charts, corporate presentation slides, and satellite imagery. In logistics or manufacturing, it could guide robots through tasks requiring visual inspection and subsequent information lookup.

The "full stack" openness—model, code, and data—is particularly significant. It allows researchers to not only use the model but also scrutinize its construction, reproduce results, and potentially identify improvements. This transparency accelerates scientific progress and allows for more rigorous benchmarking and validation of the company's performance claims, a process the broader community is likely to undertake swiftly.

Strategic Context and the Open-Source Momentum

SenseTime's move occurs within a strategic context of increasing tension and competition in global AI. While U.S. companies like OpenAI, Anthropic, and Google maintain closed ecosystems around their most advanced models, the open-source community, heavily supported by Meta and a constellation of startups, has been rapidly closing the gap. SenseTime's decision to contribute a high-performance model to the open-source camp adds considerable weight to that side of the scale.

It serves multiple strategic purposes for SenseTime: it establishes technical credibility on a global stage, attracts developer mindshare to its SenseNova ecosystem, and generates valuable feedback and downstream innovation that can inform its future commercial and proprietary products. Furthermore, it complicates the competitive landscape for U.S. giants by providing a viable, high-performance alternative that is free to use and modify.

Industry analysts note that while benchmark scores are a critical indicator, real-world utility, cost-efficiency, and ease of deployment will ultimately determine the model's impact. "The proof will be in the pudding," said the European researcher. "If developers find it genuinely more capable and reliable for building complex agentic applications than existing open-source options, it could see rapid adoption. The full开源 approach removes all barriers to trying it."

The SenseNova-MARS models and resources are now accessible via the Hugging Face repository and GitHub. The technical report detailing the methodology and benchmarks has been published on arXiv. The AI community's independent evaluation of its capabilities and the innovative applications built upon it will be the next critical chapter in this story.