Google releases Gemma 4 open models

Google just dropped Gemma 4, a family of open AI models that can run offline on phones and tiny computers with voice and vision built in—and the comment section went feral. The vibes: half victory parade, half drag race. One camp is screaming “finally!” over the Apache 2.0 license and the return of base (non-instruction) models for custom tuning. Another camp is already timing laps: devs like danielhanchen rolled out quantized versions and swear they “work really well,” while minimaxir claims the small E4B flavor beats the old 27B model across benchmarks at a fraction of the size.

The plot twist? Benchmark beef. Veteran tester jwr is itching to throw Gemma 4 at a spam filter and reminds everyone that while Gemma 3 was strong, it got eclipsed—and Qwen (another popular model family) “always had more variance,” stirring a friendly rivalry. Privacy hawks love the local voice input for translation apps, and tinkerers are giddy about turning a gaming PC into a “local-first AI server.” There’s nitpicking about model sizes (E2B/E4B for mobile, a 31B beast for “agent” tasks, a 26B mixture-in-between), and yes, jokes that your Raspberry Pi just became a mini coworker. Google’s “enterprise-grade security” line drew nods, but the crowd’s real heartbeat is simple: fast, open, and pocket-sized—now prove it on the leaderboard.