VideoJev: The New Speed Standard for Visual AI Agents

In the rapidly evolving world of artificial intelligence, visual agents—AI systems that can watch video and make decisions in real-time—are becoming essential for robotics, autonomous monitoring, and complex digital workflows. However, most current Multimodal Large Language Models (MLLMs) suffer from a significant bottleneck: they are slow. These models typically "think" by generating text word-by-word, a process known as autoregressive decoding, which is computationally expensive and often unnecessary for simple decision-making tasks.

A new research paper introduces VideoJev, a family of models designed to solve this efficiency crisis. By shifting the paradigm from "generating stories about video" to "scoring decisions against video," VideoJev offers a glimpse into a future where AI agents can react to visual stimuli with unprecedented speed and precision.

Beyond Autoregressive Bottlenecks

The core innovation of VideoJev lies in its "System One" approach to visual reasoning. Instead of asking a model to write out a long-form answer, VideoJev uses a shared scoring head to evaluate specific temporal intervals and multiple-choice options. For example, if an agent needs to know when a specific event occurred in a two-hour video, it doesn\'t describe the scene; it scores every moment to find the exact span.

This structural change allows the model to bypass the slow process of generating text tokens. The result is a unified interface that handles both temporal grounding (finding "when" something happened) and question answering (deciding "what" happened) using the same efficient mechanism.

Unprecedented Speed and Performance

The practical implications of this research are staggering. In comparative tests, VideoJev-4B and VideoJev-35B-A3B demonstrated massive improvements over traditional models. On video question-answering benchmarks, VideoJev achieved performance comparable to "Thinking" models—those that perform deep reasoning—but at speeds exceeding 100x their predecessors.

For temporal grounding, the models improved mean intersection-over-union (mIoU) scores by over 10 points in some cases, while still operating 2.8 to 3.1 times faster than their backbones. This means VideoJev is not just faster; it is more accurate at pinpointing specific moments in time, making it ideal for applications like sports highlights generation or security footage analysis.

Joint Inference: Smarter Multi-Tasking

Real-world AI agents rarely ask just one question. They often need to process multiple queries about the same video history simultaneously. VideoJev introduces "Joint Inference," a method that reuses visual computations across multiple queries in a single forward pass. This allows the model to handle complex, open-ended workflows without duplicating the heavy lifting of processing pixels.

Interestingly, the research found that training the model on both grounding and question-answering tasks simultaneously actually improved performance in both areas. The temporal supervision helped the model localize evidence better, while the question-answering data sharpened its semantic understanding.

Real-World Business Applications

For businesses, VideoJev represents a shift toward more viable AI deployments. In industries like manufacturing, where an AI might need to monitor a production line and flag specific errors instantly, the 100x speedup is the difference between a reactive system and a proactive one. Similarly, in media and entertainment, the ability to search through thousands of hours of content for specific actions with high precision and low cost could revolutionize archiving and content creation.

By providing a reusable, high-speed decision interface, VideoJev sets a new benchmark for how visual agents can be integrated into complex, real-time environments.