Gemma 4 and Qwen 3.5 quietly power AI where ChatGPT can't reach

The AI story that gets almost no airtime in the frontier-model race is the one IEEE Spectrum sketched this week: what happens when the model has to fit on a $50 board and run without a data center behind it. The piece walks through a handful of live deployments where small open-weight models are doing narrow, high-stakes work in places where ChatGPT is essentially absent.

The connectivity gap is the starting point. IEEE Spectrum cites a World Bank figure that only 0.7 percent of internet users in the world's poorest countries have used ChatGPT, versus a quarter of internet users in the most developed nations. The World Bank's Banga is quoted saying that "outside the developed world, other than maybe India and China, very few countries have that combination" the frontier services assume.

Alonge's RxScanner is the emblem: a handheld spectrometer that scans a pill with infrared light, then hands the molecular profile to an AI model with a pharmaceutical database attached. Marcelo José Rovai, involved in three more of the projects, is credited with an Arduino-based electrocardiogram running in parts of Brazil that lack access to more complex equipment, malaria-mosquito detection in a number of nations, and ant-infestation ID for a Uruguayan vineyard. A drone-based cashew disease system from Bala Murugan and colleagues at the Vellore Institute of Technology in India rounds out the list. Rovai calls small AI "the most important area in AI nowadays" and names Google DeepMind's Gemma 4, released in April, and Alibaba's Qwen 3.5 as his go-to models. The workhorse hardware includes the new Arduino UNO Q, a $50 device on a Qualcomm chipset that draws 3 watts.

The direction is captured in a line from Alonge: "the future of AI is not like one giant model, at a center. I think it's millions of small, precise models deployed at the edge, each one solving like a specific problem, a specific context." IEEE Spectrum cites the research firm Counterpoint that 45 percent of smartphones shipped this year will run generative AI, and slightly more than half will be able to run a small AI model by the end of next year.

The honest caveat is that the piece doesn't publish accuracy numbers for any of these tools, doesn't name the specific countries running the mosquito detector, and doesn't say how the deployments are funded beyond a passing note that reliable power matters and "that phone battery is not going to last forever." It also flags the dependency small-model boosters rarely mention. As Rovai puts it, "we need the big models to create these smaller models." The upside worth watching is what happens when hardware makers, open-model labs and last-mile health and agriculture teams start treating this as a shared stack rather than three unrelated efforts.