Google’s Gemma 4: Open-source AI with a license that matters
A heavy forged-iron padlock sprouting open, its shackle releasing a cascade of hundreds of brass and nickel-plated key blanks onto a dark matte slate📷 Photo by Tech&Space
- ★Apache 2.0 license replaces restrictive terms
- ★Four models, no benchmarks—just promises
- ★Meta and Mistral’s open-source playbook under pressure
Gemma 4 arrives with a licensing upgrade that actually changes the game: Apache 2.0, the same permissive terms that turned TensorFlow into a developer staple. Unlike Gemma 2’s half-open Creative ML Open RAIL license—which barred commercial use without explicit approval—this version lets companies build, modify, and monetize without Google’s blessing. For startups and enterprises tired of legal gray zones, that’s a genuine unlock.
The four new models span devices from phones to workstations, but Google’s press materials are conspicuously light on specifics. No parameter counts (though the pattern suggests 2B–27B ranges), no MLPerf benchmarks, and zero mention of inference costs or hardware requirements. ‘Most capable’ is a claim, not a data point. Even the Gemma 2 27B release included comparative scores against Llama 3—here, silence.
This isn’t just a model drop; it’s a shot across Meta’s bow. Llama 3 and Mistral’s models dominate the open-weight leaderboards, but their licensing still carries friction. Apache 2.0 removes that friction entirely—assuming the models can actually perform. The real test isn’t the license; it’s whether developers bother to deploy them over proven alternatives.
The licensing shift is real. The performance claims? Still TBD.📷 Photo by Tech&Space
The licensing shift is real. The performance claims? Still TBD.
Early GitHub and Hugging Face activity suggests cautious optimism, not hype. The community’s focus? Not performance, but permissions. Threads highlight the license’s commercial flexibility, while benchmark discussions are notably absent. That’s a red flag: when developers care more about legal terms than model quality, the product isn’t speaking for itself.
Google’s timing is no accident. The EU AI Act looms, and open-source exemptions hinge on transparency—something Apache 2.0 delivers. But let’s not confuse compliance with capability. The Gemma 2 9B underperformed against Llama 3 in real-world tasks despite Google’s benchmarks. Without third-party validation, ‘most capable’ is just a better PR strategy.
The bigger play here is ecosystem lock-in. Apache 2.0 makes Gemma 4 a safer bet for Vertex AI integrations, nudging enterprises toward Google Cloud. For competitors, it’s a forced response: either match the licensing or cede mindshare to the ‘most open’ narrative. Meta’s next move just got more complicated.