There's a question that every AI company charging you money to use their model has been quietly dreading. What happens when a world-class AI runs on your phone, for free, with no subscription, no rate limits, and no data leaving your device?
On Thursday, Google answered it.
Gemma 4 is a family of four AI models built from the same research as Google's flagship Gemini 3, which Google just made fully open-source. It’s free to use, modify, and deploy in commercial products.
Here's what those models actually do:
The E2B and E4B variants run on smartphones and Raspberry Pis. They process text, images, and audio natively, on-device, without an internet connection. The E2B is 4x faster and uses 60% less battery than previous versions
The 26B MoE runs on a standard consumer GPU, an RTX 3090 or 4090, and achieves near-frontier performance using only 3.8 billion active parameters out of 26 billion total. It ranks 6th among all open models in the world
The 31B Dense model currently ranks 3rd among all open models globally, beating models with 20x more parameters
A model Google is giving away for free, that runs on hardware anyone can buy, is outperforming models that cost hundreds of millions to build and millions more to run.
But the numbers alone aren't the real story. The real story is the license.
Every previous Gemma model came with Google's own custom license, one that enterprise legal teams flagged, compliance departments questioned, and organizations avoided because the terms could change at any time. Dozens of companies chose Llama or Mistral instead, specifically because Apache 2.0 meant no surprises.
Gemma 4 ships under Apache 2.0. No MAU limits, and restrictions on redistribution, modification, or commercial deployment.
What makes this particularly pointed is the timing. As some Chinese AI labs have begun pulling back from fully open releases for their latest models, Google is moving in the exact opposite direction. Opening up the most capable models in the Gemma family's history, under the most permissive license they've ever used, right as the open-source vs. closed-source debate is reaching a peak.
Google DeepMind CEO Demis Hassabis described them as "the best open models in the world for their respective sizes."
The AI arms race has always been about who builds the smartest model. But the more interesting battle, the one that actually determines who wins in enterprise and government, is about who makes the most capable AI the most accessible.
Google just made a very loud statement about which side of that battle it's on.
OpenAI bought the media outlet that covers it most - TBPN
The Technology Business Programming Network (TBPN) is hosted by former founders John Coogan and Jordi Hays. Sam Altman, Satya Nadella, and Mark Zuckerberg all appear on it regularly. It's on track to make $30 million this year.
OpenAI bought it this week. The show will now report to Chris Lehane, the man who ran a crypto super PAC that spent hundreds of millions on US elections, and has been whispering policy recommendations into the Trump administration ever since.
Microsoft just committed $10 billion to Japan.
The investment covers AI infrastructure, cybersecurity, and training 1 million Japanese engineers by 2030. Japan faces a projected shortfall of over 3 million AI and robotics workers by 2040. Its government has committed ¥60 trillion to science and technology investment.
And critically, Japanese companies and government agencies want AI infrastructure that keeps sensitive data inside Japan, not on American cloud servers.
A man used AI to build a $1 billion telehealth company. The secret ingredient is a weight-loss drug.
The company, Hims & Hers, combined AI-powered health consultations with GLP-1 prescriptions to create a telehealth business that crossed a $1 billion valuation this year.
AI handles patient intake, symptom analysis, and personalised treatment suggestions. Doctors review and approve. The drugs ship directly to patients. The result is a healthcare model that's faster, cheaper, and more scalable than a traditional clinic.
FAST BREAK
On a university-level math competition called AIME 2026, Gemma 4's 31B model scored 89.2%. The previous generation, Gemma 3, scored 20.8% on the same test.
Same model family, one generation apart. A jump from "can handle maths" to "better than most PhD students."
This is what a step-change in AI looks like. Not a gradual improvement but a jump so large it reads like a different product.
The models being built today are not marginally better than last year's models. They are categorically better. Anyone still making decisions based on 2025 AI capabilities is working with an outdated map.