Witbe launches its own vision-language AI models, faster and less expensive
In a press release published on November 16, 2026, Witbe presented the Witbe VLMs, a family of vision-language models designed and trained by its R&D teams and launched in September 2026.
The group, listed on Euronext Growth, highlights two advantages for these models: faster execution and a lower cost than general-purpose AI models, with the option to deploy them within its customers' infrastructures.
Models trained on Witbe's real-world test data
Vision-language models combine computer vision and natural language processing to link videos to text. Witbe's models are intended for automating video testing on all types of terminals.
They complement the general-purpose models and state-of-the-art models used by the agents of Witbe Agentic AI, the infrastructure launched in 2025 for automated testing and monitoring of video streams and streaming services.
According to Witbe, general-purpose models, trained on data freely available on the Internet, contain very little information about TV set-top box virtual keyboards, application launch grids, or image freezing on a connected TV more than five years old. Witbe VLMs are trained on these specific situations.
Yoann Hinard, Operations Director, states that specialization does not reduce generalization capability: “On applications our VLM models had never seen during training, they show performance comparable to state-of-the-art general-purpose models.”
A rollout underway, backed by AI Tokens already sold
Witbe VLMs are being rolled out. Using them requires the allocation of AI Tokens, like the existing AI capabilities of the Agentic SDK. The majority of Witbe's customers already use agentic testing through the Witbe 41 suite.
For these customers, the Tokens already acquired allow, according to the company, to run more tests, faster, without modifying existing scenarios. General-purpose models and state-of-the-art models remain available.
The validation method is unchanged: the models interpret, deterministic logic validates, and each result relies on performance indicators measured on a real device, with recorded video evidence. The AI Dial tool still allows adjusting the share of steps entrusted to AI, from 0% to 100%.
Mathieu Planche, Deputy Chief Executive Officer, indicated that customers use the Agentic SDK “in production on thousands of Witbox units deployed in the field,” and that the new models should make it possible “to offer our customers a more economical solution and to speed up scenario execution.”