UNESCO recently advocated a shift away from resource-intensive AI models in favor of more compact ones. The UNESCO report calls for governments and industry to invest in sustainable AI research and development. as a way to reduce energy consumption by up to 90%.
While much of the AI industry is still obsessed with building more data centers, on-device AI is challenging the notion that “bigger is better”.
Driven by growing demand for localized user privacy, environmental concerns à la ‘tokenmaxxing’, and efficiency required by some use cases, breakthroughs in smaller purpose-built AI models have shown developers that capable AI does not always require large cloud setups. Two trends have made this possible: smartphone chips now ship with dedicated AI accelerators, and new training techniques let much smaller models retain the quality that once required billions more parameters.
Tether AI Research recently shared the open-source release of VisionPsy-Nano, a compact vision-language model (VLM) purpose-built for on-device and edge deployment that outperforms comparable models in its weight class, including Liquid AI’s LFM2.5-VL-450M and Hugging Face’s SmolVLM2-500M.
In practice, this demonstrates how purpose-built AI models can maintain data privacy without compromising on performance. Because the model runs entirely on the device, images and questions never leave the phone, there is no upload to a server, no request log in the cloud, and the feature keeps working with no internet connection at all.
Vision language models
A vision-language model is an AI system that pairs a text-processing language model with an image encoder, enabling it to perceive and reason about visual data. In the last couple of years, even the largest cloud providers have moved into compact VLMs, with Google evolving its lightweight open-weight lineup from PaliGemma in 2024 through the Gemma 3 and Gemma 4 families.
But for AI that runs directly on a smartphone, the real frontier sits at a much smaller scale: models of around half a billion parameters, small enough to fit within the memory and power budget of an everyday phone while still delivering dependable quality. This size class has become one of the most active fronts in open model development, with Liquid AI’s LFM vision models and Hugging Face’s SmolVLM and nanoVLM series as the closest peers at this scale. Over the past several months, these compact open models have matured rapidly, with new architectures and training techniques that optimize for size and efficiency.
When AI needs to interpret physical surroundings and answer questions about them, response speed is vital. To meet different deployment needs, Tether’s vision-language model comes in two variants: VisionPsy-Nano 460M (tuned for quality) and VisionPsy-Nano-460M-Flash (tuned for latency). Despite its compact size, VisionPsy-Nano leads its weight class on several evaluation categories, including document understanding, visual perception, reasoning and knowledge, instruction following, and reliability.
Latency is the time between submitting a request to an AI application and receiving its response. This could be through an AI chatbot or voice assistant. The quicker the response, the lower the latency. VisionPsy-Nano 460M-Flash is engineered for fast, real-world deployment on everyday smartphones (e.g., cutting time-to-first response by up to 23x on a Pixel 9) while retaining around 99% of full model quality.
For the average user, a delay in feedback from an AI system might be a slight annoyance for casual tasks, but it can also cause serious problems when the AI is being used for time-sensitive tasks such as medical diagnosis, emergency response, and navigation. In a worst-case scenario, it can lead to catastrophic operational failure.
Towards embodied AI and robotics research
Beyond immediate and practical applications, such as enabling a phone to understand a document, a chart, or a scene, smaller VLMs are being explored for their potential in embodied AI, often trading off some accuracy for the efficiency needed in mobile or autonomous systems.
Applications like real-time AR, robotics, and BCI require instant responses. This is particularly true in Scene Visual Question Answering (Scene VQA), where an agent must continuously evaluate its physical surroundings to answer spatial queries.
To solve this, developers are exploring smaller VLMs for embodied AI and mobile autonomous systems. These edge systems explicitly trade off a marginal degree of general reasoning accuracy in exchange for the ultra-low latency and computational efficiency required for local execution. By executing Scene VQA directly on edge hardware, mobile systems can perceive, reason, and act in real time without depending on a cloud connection.
VisionPsy-Nano is part of Tether AI Research’s broader philosophy to bring intelligence to everyone, running entirely on device.
(Tech Crunch)