Google's shift from private AI to open-weights models

Google's shift from private AI to open-weights models
Image: Intellect (generated)

In short: For years, Google kept its AI source code under wraps. That changed with the debut of Gemma. By releasing open-weight models, Google moved toward a strategy of making AI more accessible, sparking a massive ecosystem of community-driven variants and specialized tools for everything from coding to marine biology.

Gemma was introduced in February 2024 as a lightweight, open-weight alternative to Google's proprietary Gemini models. This release marked a significant strategic pivot for the company, which had historically maintained a private stance on its AI source code. The move was widely interpreted as a direct response to competitors, such as Meta, who began releasing source code for their own AI models to foster broader developer adoption.

The evolution of the Gemma series has been rapid, moving from initial 2B and 7B parameter models to the multimodal capabilities of Gemma 3 and 4. While the standard models handle text and images, specialized variants have emerged to serve niche industries. For example, MedGemma (based on Gemma 3) is used for medical analysis, though Google clarifies it is not yet clinical grade. Other variants include CodeGemma for programming, TranslateGemma for linguistics, and ShieldGemma 2 for safety filtering of violent or explicit content.

Technically, the models utilize a decoder-only transformer architecture with grouped-query attention. A notable innovation in Gemma 3 is the separation of local and global attention, which improves the growth curve of context length costs. Furthermore, the release of Gemma 4 in April 2026 introduced a Mixture of Experts (MoE) architecture for its 26B variant. This diversity allows for high performance—with the 31B Dense variant reaching third place on Arena's text leaderboard—while remaining accessible enough for consumer devices like smartphones and tablets via the Gemma 3n series.

Opens the Intellect+ feed with this card first. Swipe up for the next one.

Sources

To cite this card, link to https://intellectplus.co.uk/app/technology-gemma-language-model.

How we source, check and correct cards: Editorial standards.

General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.

Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?

Related

Finding something worth knowing…