A new milestone achieved!
Today, Google DeepMind announced the latest achievement of its open-source Gemma "full family" suite—
For the first time, total downloads have broken through the one billion mark.
First launched in February 2024, it took two and a half years to go from 0 to 1 billion.
Now, developers worldwide have "hacked" and created 100,000 derivative variants based on Gemma, deployed it on NASA satellites, integrated it into 100 million mobile phones.......
The era of the G Emmaverse ecosystem explosion is truly here!
AI Identifies Anti-Cancer Pathways, Virtually Tests 4,000 Drugs
Now, AI has officially launched a total assault on cancer!
Among the many use cases of Gemma, what is truly shocking is that AI has begun to deeply participate in exploring new mechanisms for cancer treatment.
Just yesterday, global mRNA giant Moderna and Merck made waves worldwide by using AI to treat cancer.
In fact, as early as January this year, Google quietly entered the fray with Gemma.
A team from Google DeepMind, Google Research, and Yale University, developed a 27B parameter model based on Gemma:
Cell2Sentence-Scale 27B, abbreviated as C2S-Scale.
Paper address: https://www.biorxiv.org/content/10.1101/2025.04.14.648850v4
C2S-Scale does one thing: it writes a cell into a sentence.
For many years, cancer immunotherapy has been stuck on a significant hurdle called "cold tumors."
In simple terms, the immune system catches bad cells by relying on cells actively "holding up a sign"—
Displaying protein fragments from inside themselves on their surface, essentially telling immune cells: something is wrong with me, come check me out.
The scientific name for this action is "antigen presentation," and the structure used to hold the sign is called "MHC-I."
The problem with "cold tumors" is: they don't hold up the sign.
Immune cells pass by without seeing anything. The best immunotherapy drugs cannot hit an invisible target.
Interferon is the immune system's "alarm signal." But in real tumors, its levels are typically very low, almost non-existent.
So, the research team gave C2S-Scale a challenge:
Find a drug that does not sound the alarm itself, but can amplify a faint alarm signal in the environment to a level loud enough to make the cell hold up the sign.
This requirement of "only being effective under specific conditions" is the most difficult part of the entire experiment.
Turning a Cell into a Sentence
The smartest aspect of C2S-Scale is that it performs a very simple conversion.
Single-cell sequencing data originally appears as a large, high-dimensional vector—incomprehensible to humans and even more so to language models.
Its specific approach is to arrange gene names in a string from highest to lowest activity, with the most active at the front.
A cell becomes a sentence composed of "gene names."
With this conversion, without modifying Gemma's architecture at all, it can directly "read" cells.
Moreover, it can also incorporate pure text data like paper abstracts and cell metadata, processed for the first time by the same mechanism within the same model.
Next is the paper's core design—dual-context virtual screen.
Allow C2S-Scale to run the same experiment simultaneously in two "parallel worlds."
World A (Immunological Context Positive): Real patient-derived primary tumor samples, with a faint interferon signal in the environment.
World B (Immunological Context Neutral): Cultured cell line data, completely devoid of interferon activity.
Then, take the 4,000+ drugs and run them through these two worlds one by one.
The target is tricky: a drug that is effective only in World A, but has no effect whatsoever in World B.
Among the candidate drugs identified by C2S-Scale, only 10%-30% had been mentioned in previous literature.
The remaining 70% or more had no known association with this screening objective.
The final selection was silmitasertib (code CX-4945), a CK2 kinase inhibitor.
The model labeled it with a tag called "context split"—showing strong effects in World A, but nearly zero in World B.
In the paper, the team stated that C2S-Scale truly proposed a new, falsifiable hypothesis.
Live Cell Experiment Successfully Validates Hypothesis
A hypothesis is just that; its truth must be verified by live cell experiments.
The team used a human neuroendocrine cell model. The key was that these cell models never appeared during the model's training.
The results of the three control groups were as follows: Silmitasertib alone: no change in antigen presentation; Low-dose interferon alone: a slight increase.
Only when both were used together: antigen presentation increased by approximately 50%.
This is the first time AI has proposed a completely new mechanistic therapeutic pathway and validated it in live cells.
Google Gemma: The One Billion Milestone
Open-source AI with 1 billion downloads—that was the era of Llama.
When Google first released Gemma, the initial models were only 2B and 7B in size.
It and Gemini "come from the same school," but follow completely different paths: the flagship Gemini focuses on cutting-edge model capabilities, while Gemma is designed as a lighter, more easily deployable open model.
Developers can directly obtain the weights and run it on their own computers, mobile phones, and edge devices.
Since 2024, Gemma's growth rate has become increasingly remarkable.
In March 2025, on Gemma's first anniversary, downloads had just surpassed 100 million; by December 2025, that number had exceeded 300 million. In April 2026 with the release of Gemma 4, cumulative downloads exceeded 400 million;
By the second quarter of this year, downloads had surpassed 900 million.
Now, just over a month later, the number has officially surged past 1 billion. Simultaneously, a vast Gemma ecosystem has been born.
Beyond entering the cancer field, Google Gemma was also sent into space, transmitting its first message to Earth: "Hello, Earthlings!"
This is the first time in history that a large language model has run on high-performance GPUs in space.
Not only that, Gemma also went into the ocean and deciphered dolphin sounds.
This model is called DolphinGemma.
It directly receives dolphin vocalizations, finds recurring patterns and sequences within them, and attempts to predict: what will the next dolphin call be?
The Gemma Universe Explodes
Today, while announcing the 1 billion downloads, Google also officially launched a new GitHub repository.
Awesome Gemma, essentially the official directory for the G Emmaverse.
GitHub portal: https://github.com/google-gemma/awesome-gemma
This repository centrally collects fine-tuned models, tutorials, developer tools from the AI community, and various Gemma applications.
The large model war of the past two years could easily be understood as a competition of "whose model is the smartest."
But Gemma's 1 billion downloads reveals another dimension:
The more disruptive path for AI is to silently integrate into every piece of hardware.
References:
https://blog.google/innovation-and-ai/technology/developers-tools/gemma-one-billion-downloads/
https://github.com/google-gemma/awesome-gemma
This article is from the WeChat public account "New Zhiyuan", author: ASI Apocalypse
