Google Launches Gemma 4 — Free AI Models That Run on Surprisingly Cheap Hardware
Google DeepMind released Gemma 4 as completely free and open source — a family of four AI models that range from tiny enough for a Raspberry Pi to powerful enough to compete with the biggest names in AI, all without needing a data center.
What happened
Google DeepMind released Gemma 4 on April 2, 2026. Four AI models, completely free to use (Apache 2.0 license), and performance numbers that put them in the same league as models from companies spending billions on infrastructure.
The headline: the most powerful model in the family runs on a single high-end GPU — the kind you can rent in the cloud for a few dollars an hour. A year ago, you needed hardware costing thousands per hour to get similar results.
Four models for four different situations
Google didn't just release one model. They built four, each designed for a different use case:
- E2B (smallest) — Light enough to run on a phone or even a Raspberry Pi. Think of it as the "works anywhere" option.
- E4B (small) — A step up, meant for laptops and tablets. Still very portable.
- 26B (the clever one) — This is the most interesting model. It uses a trick called Mixture of Experts, which means it only activates a small portion of its brain for each task. The result: it's nearly as smart as the biggest model but runs about twice as fast. On AI leaderboards, it ranks #6 globally — ahead of models that need far more computing power.
- 31B (the powerhouse) — The full-sized model. Best raw performance, ideal if you want to customize it for a specific task.
All four handle text, images, and video. The two smallest ones can even process audio — something competing models at that size can't do.
It's built to actually do things, not just chat
Most AI models are designed to have conversations. Gemma 4 is designed to take actions. Google trained it specifically to call external tools and services — checking the weather, querying a database, booking something. This isn't a hack layered on top; it's baked into how the model works.
You can also toggle a "show your work" mode. For complex tasks, the model explains its reasoning step by step before acting. For simple requests, you turn that off and get faster responses. That flexibility matters when you're building real applications.
Google even shipped a demo app called Agent Skills that runs entirely on your phone — no internet needed. It can look things up on Wikipedia, create summaries, make flashcards, and chain these steps together automatically. It's a proof of concept, but it shows where things are heading: AI assistants that work offline.
You can run it from the command line
Google released a tool called litert-lm that lets you run these models with a single terminal command — no programming required:
litert-lm run --from-huggingface-repo=litert-community/gemma-4-E2B-it-litert-lm gemma-4-E2B-it.litertlm --prompt="Explain quicksort in three sentences"The smallest model runs on a Raspberry Pi 5 — a $75 computer the size of a credit card. It needs less than 1.5 GB of memory. Not blazing fast at about 8 words per second, but it works. On a modern phone chip with AI acceleration, it's much snappier.
How it compares to the competition
The open AI model space is getting crowded. Here's the short version:
- vs Meta's Llama 4: Llama can process enormously long documents (millions of words), which is unique. But it needs much beefier hardware. Gemma 4 gives you similar intelligence at a fraction of the computing cost. Also, Llama's license has restrictions for very large apps — Gemma's doesn't.
- vs Alibaba's Qwen 3.5: Very close in overall benchmarks. Qwen is slightly better at general knowledge tests, Gemma is better at math and coding. Both are fully open source.
The real story isn't which model wins on any single test. It's that Gemma 4's clever architecture lets you get 90% of the performance at a small fraction of the cost. If you're a company trying to build AI features without spending a fortune on cloud computing, that math matters.
Why this matters
A year ago, AI models this capable were locked behind expensive API calls to OpenAI or Anthropic. Now you can download one, run it on your own hardware, and never send your data to anyone's servers.
The smallest model runs on a phone. The mid-range one runs on a regular gaming PC. The biggest one runs on a single cloud GPU that costs a few dollars per hour. All completely free, no strings attached.
Google has racked up 400 million downloads of Gemma models so far. They're clearly betting that giving away the models builds a bigger ecosystem than selling access. And for developers, researchers, and companies that want control over their AI — that's genuinely good news.
If you've been wondering when free AI models would catch up to the paid ones, the gap has never been smaller.
More Articles
Anthropic Kills Third-Party Agent Access on Claude Subscriptions — Here's What Actually Happened
Anthropic's Claude Mythos Preview Can Find Zero-Days Autonomously — And They Built a Coalition to Use It
