Google DeepMind said Gemma, its open model family, has passed 1 billion downloads, roughly two and a half years after its February 2024 launch. Developers have spawned 100,000 variants, and the model has reached a NASA satellite and decoded dolphin sounds. But the most striking use is inside cancer research.
A team from Google DeepMind, Google Research and Yale built C2S-Scale, a 27-billion-parameter model that turns a single cell into a sentence of gene names ranked by activity. Without changing Gemma’s architecture, the model can then “read” cells and paper text together. They ran a dual-context virtual screen of over 4,000 drugs.
The target was a “cold tumour” that hides from the immune system by not flagging its antigens. The model hunted for a drug that works only when a faint interferon signal is present. It surfaced silmitasertib, a CK2 inhibitor, with a “context split” the literature had barely touched. Live-cell tests then showed antigen presentation rising about 50 per cent only when the drug and low-dose interferon were combined.
It is billed as the first time AI proposed a wholly new mechanistic treatment path and had it validated in living cells. Moderna and Merck, meanwhile, made their own AI-cancer headlines this week, underscoring how fast the field is moving.
Gemma’s billion downloads also mark a quieter shift: the most disruptive AI path may be the one that slips silently into every device rather than the one that wins a benchmark.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience.