IBC 2026 is a wrap! This year in Amsterdam, we unveiled our own AI models: Witbe VLMs, earned a TVBEurope Best of Show Award and a CSI Award, and took the conference stage with a peer-reviewed paper on agentic AI in production.

From agents to models: the next step

A year ago, at IBC 2025, we brought agentic testing to video services: specialized agents that plan a test, run it on a real device, and react to whatever appears on screen. At NAB 2026, that capability had grown into a full AI-native infrastructure. Today, clients run the Agentic SDK in production on thousands of Witbox units deployed in the field.

The next step was always the models themselves. So this year in Amsterdam, we introduced Witbe VLMs, our own family of Vision Language Models built for video test automation on any device.

Why train our own Vision Language Models?

General-purpose models are trained on the open web. And the open web contains very little about set-top box on-screen keyboards, launcher grids, or the way a picture freezes on a five-year-old Smart TV.

Rather than only leaning on general-purpose AI, we trained Witbe VLMs on the data we know best: video services running on real devices, in real conditions. The messy reality of testing across platforms, operating systems, and languages.

The question we heard most on the booth: does a purpose-trained model give up generality? It does not. On applications the models had never seen in training, Witbe VLMs perform on par with frontier models, with zero-shot autonomous navigation.

What changes for QA and operations teams

Faster tests

Lower latency means test suites finish faster, and device control in the REC is more responsive. Witbe VLMs run up to 2x faster per test than general-purpose models.

Lower, more predictable cost

Up to 2x cheaper per test, with more predictable behavior and more control over model upgrades. For clients already running agentic testing through Witbe Suite 41, the AI tokens already purchased now cover more testing, completed faster.

AI sovereignty

Witbe VLMs can be deployed inside the client's own infrastructure. Test data, including recorded video of the client's service, never leaves that environment. For operators facing data residency rules or content protection obligations, this is what makes AI in production possible at all.

What does not change: how a result is proven

Models interpret; deterministic logic validates. Every result stays anchored in KPIs measured on a real device, with recorded video evidence for every run.

Witbe VLMs are queried through the Agentic SDK, from the Witbox already deployed in the field. No new hardware, and existing scenarios are unchanged. Teams keep full control of the AI ratio, from 0% to 100% per scenario:

• 100% AI: full self-driving for exploratory testing and overnight UI redesign validation
• By-design AI steps: AI handles dynamic content and human-like assertions, deterministic logic measures KPIs
• Recovery and error qualification: the happy path runs algorithmically, AI activates only on unexpected events
• 0% AI: pure algorithmic tests, such as channel change and zap time measurement

Intelligence scales exactly where it earns its place, and cost stays predictable. General-purpose and frontier models remain available alongside Witbe VLMs.

On the booth: one platform, running on real devices now powered by Witbe VLMs

At Booth 1.C53, visitors saw the full AI-native platform in action, from a prompt to a running test on real Smart TVs, set-top boxes, and mobile devices:

• Witbox: real-device execution across Smart TVs, set-top boxes, mobile devices, and browsers
• REC (Remote Eye Controller) AI Assist: real-time operational control of the device fleet, including natural language interaction
• Test Studio with AI Test designer and Agentic SDK: hybrid programmatic advanced testing
• Smartgate AI: intelligent analysis that detects, explains, and prioritizes issues automatically

Real devices. Real results. Real proof, live at the booth.

Recognition: two awards, one idea

Witbe VLMs earned a TVBEurope Best of Show Award. Our AI-native infrastructure won a CSI Magazine Award 2026 for Best Monitoring or Network Management Solution, a category we also won at IBC 2025.

One for the models we trained ourselves, one for the platform they run on. Both anchored in the same idea: purpose-built AI on real devices running alongside generic model-based automation.

Together with the TV Tech Best of Show at NAB 2026, it rounds off a year where the industry kept sending the same message: AI in testing and monitoring earns its place when it proves its results on real screens with Witbe.

On stage: Agentic AI for Smart TV QA

Away from the booth, our COO Yoann Hinard presented our peer-reviewed IBC paper, "Agentic AI for Smart TV QA: A production field study across multiple platforms and languages," in the Streaming Technology 2 conference session on Saturday, September 12.

The paper covers eight months of agentic AI running in production, in an environment where applications, operating systems, and homepages change constantly, and where traditional capture-based automation becomes nearly impossible to maintain. The headline result: one engineer increased Smart TV test coverage 5x.

Real deployments, real numbers, and what it actually takes to keep tests running when interfaces refuse to stay still. Let us know if you would like to hear more about it!

What's next?

IBC 2026 is over, but the conversations are not. The team is back with a list of follow-up demos, and Witbe Suite 41 continues to prove its undeniable value within our client base.

Next stop: CABSAT 2026, October 5 to 7 at the Dubai Exhibition Centre, Expo City Dubai. Find us on Stand H2-D45, within the French Pavilion by Business France.

Thank you to everyone who stopped by Booth 1.C53, sat through a demo, and asked the hard questions. See you in Dubai!

See Witbe Suite 41 and the first AI build for video testing and monitoring: Witbe VLMs on your own setup

A demo mapped to your devices, your applications, and your stack: Book a demo