Table of Contents
Gemma Open Model for Kids: A Billion Downloads and Orbit
The Gemma open model hit a billion downloads and now runs in orbit. What open weights mean for kids who build, plus a first project and a comparison table.
A version of Gemma small enough to fit on a gaming laptop is currently flying in orbit, analyzing satellite images and deciding which ones are worth the bandwidth to send home. Google announced on August 20, 2026 that the Gemma family had passed one billion downloads, with more than 100,000 community variants published. For a Gemma open model, kids get something no chatbot subscription offers: the actual weights, on your own machine, doing what you tell it.
That distinction, open weights versus an API, is the most useful technical idea in this whole story.
Key Takeaways
- Google announced on August 20, 2026 that Gemma had surpassed one billion cumulative downloads, with over 100,000 community-published variants.
- NASA JPL flew a 4-bit compressed Gemma 3 4B on Loft Orbital’s YAM-9 satellite; the NAVI-Orbital system scored 88% accuracy on a 7,960-image benchmark and ran live captures over Toulouse, France and the Argentine coast.
- Gemma 4, released April 2, 2026, moved to the Apache 2.0 license; earlier versions used Google’s own Gemma Terms of Use.
- An open model runs offline, costs nothing per query, and cannot be switched off by a vendor or a government, which matters more after the 19-day Claude export-control shutdown in June 2026.
- The trade-off is real: open models of this size are less capable than frontier API models, and the setup work is on you.
What “open model” actually means, and what it does not
An open-weights model is one whose trained parameters you can download and run yourself. That is different from open-source software in the strict sense, because the training data and the full training code usually are not published, and different again from an API model like GPT-5.6 or Claude Opus 5, where the weights stay on the company’s servers and you rent access by the token.
Gemma is Google’s open-weights family. The timeline matters for a kid picking a version: Gemma 1 in February 2024 (2B and 7B parameters), Gemma 2 in June 2024, Gemma 3 in March 2025 (1B, 4B, 12B, 27B), and Gemma 4 on April 2, 2026 (E2B, E4B, a 26B mixture-of-experts, and a 31B dense model). Licensing shifted too. Gemma 1 through 3 shipped under Google’s own Gemma Terms of Use, which is source-available with use restrictions; Gemma 4 moved to Apache 2.0, a standard permissive open-source license. If your teen cares about what they are legally allowed to build and ship, that change is the headline, not the download count.
The billion-download announcement also lists what people built: India’s National Health Authority integrated Gemma 4 into Aarogya Setu 2.0, an app with more than 100 million Android downloads; Yale and Google researchers built C2S-Scale on Gemma and reported discovering a cancer therapy pathway later verified in living cells; DolphinGemma, with Georgia Tech and the Wild Dolphin Project, analyzes dolphin vocalizations. Over 1,600 projects were submitted to a Gemma Challenge on Kaggle.
The satellite story, told properly
The space deployment is the detail kids latch onto, so it is worth getting right rather than hand-waving.
IEEE Spectrum reported on July 23, 2026 that NASA’s Jet Propulsion Laboratory flew a 4-bit compressed version of Gemma 3 4B aboard Loft Orbital’s YAM-9 satellite. The software framework, NAVI-Orbital, was built by Juan M. Delfa and Taran Cyriac John at JPL with Andrew W. Herson at Loft Orbital, and it runs on an Nvidia Jetson Orin AGX drawing from solar panels supplying 150 to 500 watts. Benchmark accuracy was 88% classifying a 7,960-image dataset without fine-tuning on it. Two live in-orbit capture tests followed, one over Toulouse, France, and one over the coast of Argentina.
The engineering reason is the part worth explaining. Satellites have severely limited downlink bandwidth. Sending raw imagery costs megabytes; sending a text description of what is in the image costs kilobytes. Running the model onboard turns pictures into semantic summaries, so the expensive radio link carries conclusions instead of pixels. Delfa’s description of the workflow: “Now, a scientist can write a prompt, upload it to the spacecraft, and that will be taken into account by the system.”
A 4-billion-parameter model, compressed to 4 bits, on a single embedded board, in orbit. That is the concrete argument for why small models matter.
How to Teach Your Kid About Open AI Models
Ages 5–8: Borrowed versus owned
Compare a library book to a book you own. Both let you read. Only one lets you keep it forever, take it camping with no internet, and write notes in the margins. API models are the library book; open models are the one you own. Then ask which they would take on a camping trip, and why the library is still useful.
Ages 9–12: Run one and pull the plug
With a parent, install Ollama or LM Studio on a family computer and download a small Gemma model (the 1B or 4B size runs on ordinary hardware). Ask it a few questions. Now turn off the Wi-Fi and ask it more. The moment a kid sees an AI answering with the internet off is the moment “it lives in the cloud” stops being their mental model. Follow with the honest comparison: ask the same question to a frontier chatbot and let them see the quality gap.
Ages 13+: Build something that could not exist on an API
Pick a project where offline or private operation is the point: a notes summarizer that never sends schoolwork to a server, a classifier for photos of a science experiment, a chatbot that only knows a specific set of documents. The constraint is the assignment. Then have them read the license of whichever Gemma version they used and write down, in their own words, what they are allowed to do with what they built. Reading a license is a genuine engineering skill and almost nobody teaches it.
The question to ask: “The satellite version of Gemma had to fit on one small board with limited power. What did the engineers give up to make it fit, and what did they gain?”
Open models versus API models: the honest comparison
| Open model (Gemma) | API model (GPT-5.6, Claude Opus 5, Gemini) | |
|---|---|---|
| Where it runs | Your laptop, a Raspberry Pi, a satellite | The company’s data center |
| Cost per question | $0 after download | Per token; Opus 5 is $5/$25 per million |
| Works offline | Yes | No |
| Can be switched off by someone else | No | Yes: Claude Fable 5 was pulled worldwide for 19 days in June 2026 |
| Capability | Good for its size; well below frontier | Frontier |
| Your data leaves your machine | No | Yes, subject to the provider’s retention policy |
| Setup difficulty | Moderate: install a runner, pick a size, manage RAM | None: open an app |
| License to build on | Apache 2.0 for Gemma 4; Gemma Terms of Use for 1 through 3 | Governed by commercial terms; model copying is restricted |
| Good first project | Offline summarizer, image classifier, document Q&A | Anything needing top reasoning quality |
The row about being switched off is not hypothetical. On June 12, 2026, a U.S. Commerce directive required Anthropic to bar foreign nationals from Claude Fable 5 and Mythos 5; unable to verify nationality in real time, the company disabled both worldwide until controls were lifted June 30. Anyone who had built on the API had nothing for 19 days. Anyone running weights on their own machine was unaffected. Our explainer on AI export restrictions covers that episode.
Be honest about the capability column with your kid, though. A 4B open model is not going to match Claude Opus 5 on a hard reasoning problem, and pretending otherwise sets them up for disappointment.
Why a Gemma open model matters for kids who build
Three reasons, in order of how much they actually matter.
It turns AI from a service into a component. A kid who has downloaded weights and wired them into a Python script has a different relationship with AI than one who only types into a chat box. They can now ask “what could I build with this” instead of “what can this do for me.” That shift is the whole point.
It removes the cost ceiling on experimentation. Zero marginal cost per query means a teen can run a thousand experiments without a parent watching a bill. With API pricing, ambitious experimentation gets expensive fast.
It teaches constraints, which is what engineering is. The satellite team compressed a model to 4 bits to fit 8 GB and a power budget. Choosing a model size against available RAM is the same problem in miniature, and it is the kind of decision that professional ML engineers make constantly.
The education-outcome evidence does not extend to any of this, and it would be dishonest to claim otherwise. No study shows that running a local model improves grades. What it improves is the category of thing a kid believes they can make, which is harder to measure and, in my experience building consumer hardware, more predictive of who keeps going. For related paths, see our piece on open-source versus proprietary AI.
What to Watch For Over the Next 3 Months
- Week 4: Install one small open model on a family computer and run it with the Wi-Fi off. Fifteen minutes, one memorable moment.
- Month 2 red flags: A teen downloading models with no idea what the license permits; a project that quietly sends private data to a server they did not choose.
- Month 3 self-check: Ask your kid to name one thing their local model does better than a chatbot subscription (privacy, offline, cost) and one thing it does worse (raw capability). If they can do both, they understand the trade.
Frequently Asked Questions
What is the Gemma open model?
Gemma is Google’s family of open-weights AI models, meaning you can download the trained parameters and run them on your own hardware. Google announced on August 20, 2026 that the family had passed one billion downloads with over 100,000 community variants.
Can my kid actually run Gemma at home?
Yes, with a parent’s help. The smaller sizes (1B and 4B parameters) run on ordinary laptops using a tool like Ollama or LM Studio. The satellite deployment used a 4-bit compressed Gemma 3 4B on a single embedded board, which gives a sense of how small these can get.
Is Gemma really running in space?
Yes. IEEE Spectrum reported on July 23, 2026 that NASA JPL flew a 4-bit compressed Gemma 3 4B on Loft Orbital’s YAM-9 satellite as part of the NAVI-Orbital framework, with 88% accuracy on a 7,960-image benchmark and live captures over France and Argentina.
Is Gemma open source?
Gemma 4, released April 2, 2026, uses the Apache 2.0 license, which is standard permissive open source. Gemma 1 through 3 used Google’s own Gemma Terms of Use, which is source-available with use restrictions. Training data is not published for any version.
Is an open model as good as ChatGPT or Claude?
No, not at these sizes. A 4B or 27B open model is well below frontier API models like Claude Opus 5 on hard reasoning. What it offers instead is offline operation, zero per-query cost, privacy, and the ability to build on it without depending on a vendor.
Why would anyone use a small model when big ones exist?
Because of constraints: no internet, limited power, private data, or a budget of zero. The satellite case is the clearest illustration; downlink bandwidth made sending a text summary from a small onboard model better than sending raw images to a bigger model on the ground.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Google. (2026, August 20). “Inside the Gemmaverse: Celebrating one billion Gemma downloads.” The Keyword. https://blog.google/innovation-and-ai/technology/developers-tools/gemma-one-billion-downloads/
- IEEE Spectrum. (2026, July 23). “NASA Puts Google’s Gemma Large Language Model in Orbit.” https://spectrum.ieee.org/nasa-ai-satellite-image-analysis
- Wikipedia. (2026). “Gemma (language model).” https://en.wikipedia.org/wiki/Gemma_(language_model)
- Google. (2026, August). “Google AI announcements from August 2026.” https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
- Anthropic. (2026, June 12). “Statement on the directive to suspend Fable 5 access.” https://www.anthropic.com/news/fable-mythos-access
- Anthropic. (2026, July 1). “Redeploying Claude Fable 5.” https://www.anthropic.com/news/redeploying-fable-5
- Anthropic. (2026, July 24). “Introducing Claude Opus 5.” https://www.anthropic.com/news/claude-opus-5
- Unite.AI. (2026). “Google’s Gemma Open Models Pass 1 Billion Downloads as Variants Top 100K.” https://www.unite.ai/googles-gemma-open-models-pass-1-billion-downloads-as-variants-top-100k/
