Google Should Open Source Gemini. All of It. - Thomas Unise

Press enter to begin your search<br>Search

Close Search

AI<br>Google Should Open Source Gemini. All of It.

By Thomas UniseJuly 9, 2026No Comments

Google gave the world the transformer. China is now giving the world the models.

That is the problem.

In 2017, Google researchers published “Attention Is All You Need,” the paper that introduced the Transformer architecture and became the foundation for modern large language models. ChatGPT, Claude, Gemini, Qwen, DeepSeek, GLM, Kimi all live in the world Google helped create.

But Google no longer controls the direction of that world.

It invented the substrate, then watched the AI industry turn that gift into companies, products, developer ecosystems, and trillion-dollar expectations. Now the same pattern is repeating with open-weight models. Only this time, the center of gravity is not Mountain View. It is moving to China.

And that’s why Google should open source Gemini. Like, today.

Not Gemma. Not a small edge model. Not a polite developer demo that runs on a laptop and lets Google say it participates in open AI. The actual flagship weights. The model family that would instantly become the American open-weight default if Google had the nerve to release it.

This is not a charity move either, it’s a real strategy.

Google is already fighting a brutal (and losing) closed-model war against OpenAI and Anthropic.

In that arena, Gemini has to win on product polish, benchmarks, distribution, developer love, enterprise trust, pricing, speed, reasoning, coding, agents, and brand perception at the same time. That is a hard war to win when the market increasingly sees frontier AI as a rotating leaderboard.

The closed API race is also unforgiving. If Gemini slips, everyone notices.

At Google I/O 2026, Sundar Pichai said Gemini 3.5 Pro was coming “next month.” Google later reportedly pushed the release target into July while it gathered feedback and tuned the model.

That delay is not proof that Google is incompetent. It may be the opposite. It suggests a team refusing to ship a model before it is ready. But that is exactly my point.

Closed frontier AI forces Google to play the same product-release game as everyone else. Every missed window, every benchmark, and every delay becomes a story and evidence that Google is somehow falling behind in the very field it helped invent.

Open weights would change the game.

As a closed product, Gemini is one more model in a crowded API market. As an open-weight release, Gemini becomes infrastructure.

Developers do not merely test it. They build around it. Enterprises do not merely call it. They deploy it, fine-tune it, audit it, host it, wrap it, and make it part of their internal stack.

That is how defaults are created.

Google knows this better than anyone. Android, Chromium, Kubernetes, and TensorFlow are some of Google’s most important wins that came from making technology too useful and too available to ignore. The transformer paper was probably the most valuable developer marketing Google ever produced, and it did not even look like marketing. It looked like research. That was the magic.

They should do it again, but on purpose.

The case for this is even stronger because the open-weight market is not waiting for America to organize itself.

Z.ai released GLM-5.2 in June 2026, and Artificial Analysis ranked it as the leading open-weight model on its Intelligence Index shortly after launch. Z.ai described the model as built for long-horizon tasks, while outside observers highlighted its MIT-licensed open-weight release and serious coding performance.

Reuters has reported that Chinese open-source models are being widely adopted globally because of their cost and technical strength, and that Beijing is now considering restrictions around access to some AI technologies. That should terrify anyone who cares about American AI leadership.

The West spent years assuming open models were the low-end of the market. Then Chinese labs used them to win developer mindshare, enterprise experimentation, and global distribution.

The economics are obvious. Most enterprise AI work is not Nobel-level reasoning. It is document review, extraction, search, coding assistance, email triage, workflow automation, customer support, analytics, summarization, and internal agents.

For that work, “good enough and cheap enough” beats “slightly better and far more expensive” all day.

That is where open weights become a winner. Once a company standardizes on a model family, the model becomes more than a tool. It becomes a platform decision. Fine-tunes accumulate. Eval harnesses get built. Internal prompts get written. Security reviews get approved. Engineers become familiar with its quirks. Switching costs appear.

If American companies standardize on Qwen, DeepSeek, Kimi, GLM, or whatever Chinese model wins the next open leaderboard, that is not just a pricing story. It is a platform...