AI is coming to you

For a long time, the rule in the AI industry was simple: bigger model, bigger cloud, better results. In recent months, however, we’ve been seeing the opposite trend. More and more companies are turning to compact models that run directly on the phone, without a single byte of data leaving the device. For businesses considering a new website, app, or digital product with AI functionality, this is already a direction worth keeping an eye on.

Why Is AI Moving in the Opposite Direction?

Recently, UNESCO published a report recommending that the industry move toward more compact AI models. According to the report, such a shift could reduce energy consumption by up to 90%. This comes at a time when much of the AI sector continues to focus primarily on investing in new data centers.

At the same time, interest in on-device privacy is growing. Users want AI features that don’t send photos, documents, or personal data to a third-party server. Two factors are coming together here: smartphones now have built-in AI chips, while new training techniques are allowing much smaller models to achieve a level of quality that, until recently, required billions of parameters.

Less Computing Power, the Same Quality

One example is the recently released open model VisionPsy-Nano from Tether AI Research - a compact vision-language model (a system that combines text-based AI with image recognition) designed specifically to run directly on the device. With just 460 million parameters - far smaller than competing models such as Liquid AI’s LFM2.5-VL or Hugging Face’s SmolVLM2 - the model delivers better results in document understanding, visual perception, and instruction following.

The model also has a faster version, optimized specifically for response speed. It retains around 99% of the quality of the full version while responding several times faster.

Response Speed Makes All the Difference

For applications where AI needs to “understand” what it sees - navigation, real-time assistance, or scene recognition through a camera - response time isn’t simply a matter of convenience; it’s a prerequisite for the product to work at all. The faster version of the model reduces the time to the first response by up to 23 times on a tested Pixel 9. That can be the difference between an AI assistant that feels instantaneous and an app that feels like it’s “thinking” for too long.

That’s precisely why compact models are increasingly being tested beyond smartphones as well - in robotics, augmented reality, and interfaces where a device needs to continuously assess its surroundings without relying on a cloud connection.

What Does This Mean for Businesses Building Digital Products?

For a business owner considering a new website or app with an AI component - such as a chatbot, image recognition, or personalized recommendations - this trend is worth taking into account as early as the planning stage. A feature that runs locally on the device generally means faster responses for the user, less dependence on external AI providers, and data privacy.

If you’re considering a digital product with an AI component and want to assess which approach makes sense for your specific case, we’d be happy to discuss it with you.