Google's Gemma 4 Shipped Apache 2.0: Why Open Weights Just Got Serious

On April 2, Google released Gemma 4, the latest generation of its open-weight model family. The release is notable for three reasons. The license (Apache 2.0, genuinely permissive). The capability ladder (four model sizes, covering everything from phones to data centers). And the positioning. These models are pitched explicitly at agentic workflows, not just chat.

Here is what shipped, and why I think it matters for anyone actually building production AI.

The Lineup

Four variants, same family:

• E2B (Effective 2B). Phones, Raspberry Pi.
• E4B (Effective 4B). Laptops, edge devices.
• 26B MoE. Consumer GPUs.
• 31B Dense. Cloud, data-center inference.

All four ship with 256K context windows, native vision, native audio, and fluency in 140+ languages. Function calling, structured output, and code generation are built in, not bolted on. The 31B model reportedly beats several closed 400B-parameter competitors on reasoning benchmarks. Fair warning, Google's benchmark selection is always a little generous.

Apache 2.0 means you can use Gemma 4 commercially without revenue caps, distribution restrictions, or the "ethics" riders that other "open" models have quietly included. It is the cleanest license from a major lab in recent memory.

Why This Changes the Calculation

For the last two years, the default answer to "which model should I use" has been "whichever closed frontier model is cheapest per task." Open models were for hobby projects, privacy-sensitive workloads, and the thin slice of teams willing to run their own inference.

Gemma 4 narrows that gap noticeably. A 31B dense model with real agentic capability, running on one good GPU, at zero per-token cost, changes the math for a few specific use cases:

• Privacy-sensitive enterprise work where data cannot leave the VPC.
• High-volume, predictable workloads where per-token pricing dominates infra cost.
• Edge and on-device inference where network latency kills UX.
• Fine-tuned verticals where a specialized smaller model beats a general larger one.

My Unbiased Take

Gemma 4 is not going to replace Claude Opus 4.7 or GPT-5.4 for frontier coding work. It is not even close. If you looked at the benchmarks and thought "great, I can drop Opus," you are going to be disappointed.

What Gemma 4 does is lift the floor of competent AI. Before, running your own model meant accepting a meaningful quality gap for the privacy and cost benefits. Now the gap is narrower, and for a lot of real-world workloads, it is effectively gone. If you are running a customer-support agent, a document classifier, a first-draft summarizer, or a voice pipeline, the question is no longer "can an open model do this?" It is "why am I still paying closed-model prices for this?"

The other signal worth naming. Google shipping at Apache 2.0 while Meta's Llama licensing keeps getting stickier is a genuine shift. The open-weights leadership position is changing hands in real time, and the teams that benefit most are the ones paying attention to license terms, not just leaderboards.

If you are an AI engineer, download Gemma 4 this week. Spin it up locally. Benchmark it on your own workload, not someone else's. The teams that internalize the actual quality curve of open models in April 2026 will build a very different stack in 2027 than the teams that do not.