


VLX — Edge-native multimodal AI for real-time perception, reasoning, and action. GitHub: github.com/om-ai-lab Discrod: discord.gg/SEVNjyXPef


Describe it. See the results. From traffic monitoring to production lines and complex real-world environments, VLX-Seek makes it easy to locate, distinguish, and count the objects that matter. Just describe what you want, and VLX-Seek gets to work. One model, countless scenes, endless possibilities. See VLX-Seek in action. ↓

🚀 Fast-moving objects. Precise localization. On a high-speed production line, bottles move fast. Tracking and locating each one accurately can be challenging. With VLX-Seek, simply give it an instruction: "Locate all the bottled water on the conveyor belt." It identifies the bottles in the video and pinpoints their locations with bounding boxes — even as they move rapidly along the line. 🎯 Precise visual grounding, built for dynamic real-world scenes. Explore VLX-Seek 👇 [platform.om-ai.com/]

🚨 What if AI could spot emergency lane violations the moment they happen? With VLX-Flow 1.5, just set one instruction: “Watch the emergency lane and alert me if a vehicle occupies it.” The model continuously monitors the live video, identifies potential violations, and proactively alerts you when an event is detected. No need to repeatedly check the footage or ask the model what's happening. Always watching. Alerting when it matters.

See the full results and evaluation setup: om-ai-lab.github.io/2026_11_…

Try VLX flow 1.5 here platform.om-ai.com/

VLX-Flow 1.5 scores 85.00 on StreamingBench. Across three public benchmarks, we tested how well VLX-Flow 1.5 understands video as it streams, recalls what it has seen, and answers questions in real time. 🎬 StreamingBench · Real-Time Visual 85.00 overall, the highest among the models compared in our report, ahead of AURA (83.20) and MOSS-VL-Realtime (82.90). VLX-Flow 1.5 also scored 94.70 on Text Recognition and 91.83 on Object Perception. ⏪ OVO-Bench · Online Video Understanding 70.51 overall, with 78.39 on Backward Tracing, the highest among the models compared. This task tests how well a model recalls what it has already seen in the stream. 💬 TraceBench v1.1 · Real-Time Q&A 76.5% QA accuracy, the highest among the models compared, with 100% QA completion and 560 ms response latency on the Q&A track. Each benchmark examines a different aspect of streaming intelligence, from understanding ongoing scenes to answering questions and responding to events. These results provide a broader picture of VLX-Flow 1.5 beyond isolated video demos.




Smarter Livestock Farming Starts with Better Visual Understanding. 🐄 Managing a herd means keeping track of every animal, every movement, and every change. With VLX-Seek, cameras can do more than simply capture footage. They can identify cattle across pastures, locate individual animals, and count the herd through natural-language instructions. No need to manually inspect every frame or configure fixed detection zones. Just tell VLX what you're looking for. From herd counting to livestock monitoring, VLX brings smarter visual perception to modern farming. See more. Understand better. Farm smarter.

Count Every Chicken. Miss None. 🐔 Managing a poultry farm means keeping track of every bird, but manually counting hundreds or thousands of chickens is time-consuming and error-prone. With VLX-Seek 1.5, simply describe what you're looking for. The model can identify chickens across the entire farm, locate each individual bird, and count them automatically. Need to check a specific area or recount the flock? Just ask. See every chicken. Count with confidence. #VLX #VLXSeek #PhysicalAI #EdgeAI #OpenSourceAI

⏱️ In real-time video understanding, timing is part of intelligence. A model can generate the right answer and still get the interaction wrong if it responds too early, too late, or too often. VLX-Flow 1.5 first learns when to respond and when to keep watching through streaming behavior supervision. It then builds on that foundation with Streaming Reinforcement Learning, refining its response timing across full interactions with a video stream. The training process focuses on three dimensions: 1/ Semantic Quality Does the response accurately reflect the visual evidence? 2/ Temporal Reward Does the model respond within the appropriate event window, rather than jumping to conclusions or responding too late? 3/ Trajectory Comparison How does one sequence of responses compare with alternative interaction trajectories over the same video stream? The model learns to distinguish between: Responding before an event occurs. Responding at the right moment. Repeating an event that has already been reported. Remaining silent when no response is needed. The objective isn't just to generate the right answer. It's to respond at the right moment — and stay silent when nothing needs to be said.

Try VLX Flow 1.5 here platform.om-ai.com/

🎥 Real-time video understanding, as the world unfolds. VLX-Flow 1.5 continuously processes live video streams, understanding what's happening as scenes evolve. ⚡ What makes it different? • Continuous Perception — Processes incoming video without repeatedly analyzing the entire history. • Persistent Context — Remembers what happened earlier and connects it to what's happening now. • Real-Time Understanding — Tracks changes, understands ongoing activities, and delivers context-aware responses. Built for continuous video understanding on edge devices. Watch VLX-Flow 1.5 in action. 👇 #VLX #VLXFlow #EdgeAI #PhysicalA

What happens when a video stream never stops? Reprocessing the entire history every time a new segment arrives isn't a practical solution. VLX-Flow 1.5 uses an incremental streaming architecture to carry useful information forward as new video arrives. 🔍 Three components make this possible: 1. ⚡ Incremental Visual Encoding New video segments are processed as they arrive, avoiding repeated encoding of the full historical video. 2. 🧠 Dual-Layer Memory Visual Cache preserves recent visual details. Semantic Memory maintains useful context from previous observations and interactions. 3. 🔄 Recurrent Linear Attention Recurrent states carry historical information forward, allowing the model to continue reasoning without revisiting the entire video history. 🚀 The result is a more efficient approach to long-running video inference, with first-token latency and memory usage remaining stable over time. Continuous perception. Persistent context. Incremental computation.