Every few months, a model release comes along that is less about the benchmarks and more about what the benchmarks are now free to run on. On April 2, 2026, Google DeepMind released Gemma 4 — and the headline is not that the 31B model scores 89.2% on AIME 2026 or that its Codeforces ELO jumped from 110 to 2,150. The headline is the license: Apache 2.0, the first fully open-source release in the Gemma family's history.
Previous Gemma models carried usage restrictions that limited commercial deployment and required compliance with a custom acceptable-use policy. Apache 2.0 removes all of that. No monthly active user caps. No acceptable-use restrictions. No lawyers required before shipping a product. Developers can now download Gemma 4, build on it, and deploy it commercially without asking Google's permission — full stop.
What Is Gemma 4 and Why Does the License Matter
Gemma 4 is a family of four models designed to cover the full spectrum of inference environments — from devices with limited memory to cloud infrastructure running full-scale workloads. The lineup includes a 5B dense model, an 8B dense model, a 26B mixture-of-experts model with just 4B active parameters per token, and a 31B dense model aimed at maximum capability.
The 26B MoE architecture is particularly notable: it delivers near-31B quality at a fraction of the compute cost, because only 4 billion parameters are active at any given inference step. For teams trying to run capable models on constrained hardware, this is the model to watch.
But the license change matters more than any single architecture decision. Meta's Llama series has dominated the open-source model ecosystem partly because of its permissive licensing relative to alternatives — but Llama's license is still a custom agreement with its own restrictions. Apache 2.0 is an OSI-approved, universally recognized open-source license that enterprise legal teams already understand and have approved. Google just made the procurement conversation at enterprise companies considerably shorter.
Benchmark Performance: Where Gemma 4 Stands
The 31B model's benchmark numbers are strong across the board. On AIME 2026 — the math competition benchmark that has become a proxy for frontier reasoning capability — it scores 89.2%, compared to 20.8% for Gemma 3 27B. That is not an incremental improvement; it is a generational jump.
On GPQA Diamond, the science reasoning benchmark that tests graduate-level physics, chemistry, and biology, Gemma 4 31B scores 84.3%. LiveCodeBench v6, which measures real-world coding performance on problems that postdate training cutoffs, comes in at 80.0%. MMLU Pro hits 85.2%. On the LMArena text leaderboard, it ranks #3 globally among open models, with an ELO of approximately 1,452.
For context: the Codeforces ELO improvement from 110 to 2,150 means the model crossed from near-novice to expert-competitive programmer territory in one generation. That is the kind of capability jump that changes what open models can be used for in production environments.
Four Model Sizes, One Strategy
Google's four-model release strategy reflects a realistic view of where AI is actually deployed. The 5B and 8B models are designed for edge and on-device inference — fast, low-memory, capable of running on consumer hardware or mobile chips without a network connection. Both support context windows up to 128K tokens. Both include native multimodality: image, video, variable-resolution visual inputs, with the E2B and E4B variants adding native audio input as well.
All four models are natively trained on 140+ languages, which is not a fine-tuning afterthought but a core training objective. For developers building multilingual products, this matters. Most frontier models remain English-dominant in practice regardless of what their documentation claims; Gemma 4's training signal is structured to take multilingual performance seriously from the start.
Available Everywhere, Immediately
Unlike some model releases that announce availability in stages, Gemma 4 shipped on day one across the major deployment surfaces: Google AI Studio, Hugging Face, Kaggle, Ollama, and Google Cloud. Developers can pull the weights on Hugging Face immediately and run local inference via Ollama with a single command. Google AI Studio provides hosted access for those who do not want to manage infrastructure.
The breadth of immediate availability is itself a signal. Google is not trickling out access to manage demand or protect competitive position. They are releasing broadly and quickly, which is how you compete for developer trust in a market where Meta's Llama has built a massive ecosystem through exactly this kind of open distribution.
What This Means For You
If you are a developer, the 5B and 8B models are worth experimenting with today. Pull them from Hugging Face or run them locally with Ollama. The 26B MoE is the most interesting architecture for teams that want near-top performance without full 31B inference costs. All four support 128K context minimum.
If you work at a company evaluating open-source AI for production, the Apache 2.0 license removes the main legal objection your procurement team has had. This is a model your legal counsel will approve without a custom review. That friction was real and it is now gone.
If you are following the frontier model race: Google just moved decisively into the open-source developer ecosystem with a model that matches or beats the previous generation of closed competitors on key benchmarks. The gap between what you can run locally and what you can access only through an API is narrowing. Gemma 4 is evidence that the narrowing is accelerating.