A 27 billion parameter model that answers questions in more than 140 languages, reads a 128,000 token document in one pass, and still fits on a single consumer GPU sounds like a niche curiosity. It isn’t. It’s Gemma 3, released by Google in 2025, and for most of its life the technical part of that description mattered less to enterprise buyers than four words buried in a licensing document: “custom, non-Apache license.” That detail, not any benchmark, is the real story of the Gemma family, and it only got resolved in April 2026.

From Gemini’s research line to a laptop-sized model

Gemma 1 arrived in February 2024 as a deliberately scaled-down sibling of Google’s Gemini models, built from the same research and training techniques but sized for people without a data center. Gemma 2 followed later that year, refining the same idea: smaller architectures that behaved like larger ones on reasoning and instruction-following benchmarks. Neither release changed the plan much, they widened it. By the time Gemma 3 shipped in 2025, the lineup spanned 1B to 27B parameters, added multimodal input (text and images together) starting at the 4B tier, extended context to 128,000 tokens, and covered more than 140 languages. Every one of those sizes was designed to run on a single GPU, which is a different design goal than running well in a cluster. It meant a researcher, a small company, or a hobbyist could load a capable model onto one card and get real work done without provisioning a rack.

What changed in April 2026

Gemma 4, released in April 2026, kept extending that same trajectory on the technical side: hybrid attention pushed the context window to 256,000 tokens, and the smallest variants became efficient enough to run directly on a phone. Those are genuine engineering gains. But the change that mattered most to the people deciding whether to build on Gemma wasn’t in the model card, it was in the license file. For its first three generations, Gemma shipped under a custom license that was more restrictive than the Apache 2.0 or MIT terms most open-source infrastructure runs on, with extra conditions on redistribution and derivative use, the kind of clauses that make a legal team pause before a procurement request even reaches an engineer. With Gemma 4, Google switched to Apache 2.0 outright.

Why the license mattered more than the benchmarks

Picture a car that is fast, reliable, and cheap to maintain, sold only under a lease that forbids you from driving it across state lines or ever selling it to someone else. None of that changes how well the engine runs. It changes whether a rational buyer signs the lease. That was Gemma for two years: a genuinely capable model, competitive on the benchmarks people cared about, that a lot of legal and procurement teams wouldn’t clear for production because of terms that had nothing to do with what the model could do. The single-GPU footprint, the 140-language coverage, the long context window, none of it mattered to a compliance review that stalled at “custom license, pending legal.”

The move to Apache 2.0 in April 2026 is worth paying attention to for reasons beyond Gemma itself. It’s a data point about the whole open-model space: adoption tracks licensing risk at least as closely as it tracks capability. A model that’s slightly behind on benchmarks but cleanly licensed will often get deployed faster and more widely than one that leads every leaderboard but carries ambiguous redistribution terms. Gemma spent three generations proving the reverse case, strong technically, held back contractually, before its license caught up with its engineering.

For a closer look at why that distinction, open source versus merely open weight, keeps deciding real adoption outcomes, see Open Source vs Open Weight. And since Gemma’s whole design philosophy has been built around fitting onto modest hardware, from a single GPU down to a phone, Hardware for Open Models is the piece that lays out the hardware tiers that determine what you can actually run.