Google Releases Gemma 4 Under Apache 2.0, Dropping Its Custom AI License

Google DeepMind released Gemma 4 on Thursday, a family of four open models built from the same research that powers its proprietary Gemini 3, the company announced. The bigger news sits in the fine print. For the first time, Google is shipping its open models under a standard Apache 2.0 license, abandoning the restrictive custom terms that kept enterprise legal teams nervous for two years.

The shift matters because Google's previous Gemma license included a prohibited-use policy the company could update unilaterally. It required developers to enforce Google's rules across all derivative projects and, according to Ars Technica, could even be read to transfer licensing obligations to AI models trained on Gemma-generated synthetic data. Apache 2.0 eliminates all of that. No custom clauses, no redistribution restrictions, no commercial carve-outs.

Four models, two deployment tiers

Gemma 4 ships in four sizes organized around where they run. The 31-billion-parameter dense model and its 26-billion-parameter MoE sibling target workstations, handling text and images across 256,000-token context windows. Then the small ones. E2B and E4B were built for phones, Raspberry Pi boards, Jetson Orin Nano modules. They get 128,000-token context and native audio, so speech recognition runs entirely on the handset.

The naming convention needs unpacking. VentureBeat reports that the "E" prefix denotes "effective parameters." E2B activates 2.3 billion effective parameters during inference but carries 5.1 billion total, because each decoder layer includes its own embedding table through a technique called Per-Layer Embeddings. The architecture is large on disk but cheap to compute.

Inside the 26B MoE model, 128 small experts sit waiting. Eight activate per token, plus one shared always-on expert. Only 3.8 billion of the model's 25.2 billion total parameters fire during inference. That translates to 26B-class intelligence at roughly 4B-class compute costs, according to Google.

Benchmarks tell a mixed story

Google positioned the 31B dense model at third place among open models on the Arena AI text leaderboard. But the rankings shift depending on which benchmark you read.

On AIME 2026, the American Invitational Mathematics Examination adapted for AI evaluation, the 31B model scored 89.2% and hit a Codeforces ELO of 2,150. Those numbers would have been frontier-class from closed models not long ago. Gemma 3 27B, by comparison, managed just 20.8% on AIME without thinking mode enabled. The generational leap is real.

Independent testing from Artificial Analysis adds context you won't find in Google's press materials. On GPQA Diamond, a graduate-level science reasoning benchmark, Gemma 4 31B scored 85.7% in reasoning mode. That's second-best among all open models under 40 billion parameters. Alibaba's Qwen 3.5 27B scored 85.8%. The gap is a rounding error, but it means Google's flagship open model doesn't hold the crown. Qwen 3.5 also beats Gemma 4 on MMLU-Pro (86.1% vs. 85.2%) and Humanity's Last Exam, where Chinese models dominate.

Where Gemma 4 claws back ground is efficiency. At around 1.2 million output tokens, the 31B model reportedly consumes less compute than Qwen 3.5 27B (1.5 million) or Qwen 3.5 35B (1.6 million). For teams paying per GPU-hour, that spread adds up.

The license is the real product

For two years, enterprises evaluating open-weight models faced an awkward choice. Google's Gemma line performed well, but legal departments flagged the custom license. Compliance teams hesitated. And many organizations chose Mistral or Alibaba's Qwen instead, because those shipped under terms lawyers already understood.

Hugging Face CEO Clement Delangue did not hedge. "A huge milestone," he called the Apache 2.0 switch. Demis Hassabis went further, branding Gemma 4 "the best open models in the world for their respective sizes." That's the Google DeepMind CEO selling hard, but the licensing shift gives the claim teeth.

The timing carries its own signal. Chinese AI labs have started pulling back from fully open releases, VentureBeat reports. Alibaba's latest Qwen models, including Qwen 3.5 Omni and Qwen 3.6 Plus, arrived with tighter restrictions. Google is moving in the opposite direction, opening up while competitors close down.

What actually ships on your phone

Google confirmed to Ars Technica that Gemini Nano 4, the on-device model running on Pixel phones, will be based directly on Gemma 4's E2B and E4B. That's the first official confirmation of the next Nano generation. Developers can prototype agentic workflows in the AICore Developer Preview now, with forward compatibility promised when Nano 4 launches later this year.

The edge models punch above expectations. E4B hit 42.5% on AIME 2026 and 52.0% on LiveCodeBench, strong numbers for a model that runs on a T4 GPU. Google's audio encoder shrank from 681 million parameters to 305 million, and frame duration dropped from 160 milliseconds to 40, making real-time transcription viable on a handset.

Function calling is baked in at the architecture level across all four models, drawing on research from Google's FunctionGemma release late last year. Earlier open models faked it. Developers would write elaborate prompts hoping the model cooperated with structured tool calls. Hit or miss. Gemma 4 trains the capability from scratch.

The ecosystem play

Developers have pulled down Google's open models four hundred million times since February 2024, spinning off more than 100,000 community variants along the way. That installed base existed before this release. Some of the stranger spinoffs hint at the range: MedGemma reads medical imaging, DolphinGemma parses dolphin vocalizations, and SignGemma handles sign language translation. Not the obvious use cases, which is the point.

NVIDIA announced day-one optimization for Gemma 4 across its GPU lineup, from Jetson Orin Nano to Blackwell. The models work with Ollama, llama.cpp, MLX, LM Studio, vLLM, and Unsloth out of the box. Google Cloud offers serverless deployment through Cloud Run with RTX Pro 6000 GPUs that scale to zero when idle.

Holger Mueller at Constellation Research framed it as land-and-expand: "Google is building its lead in AI, not only by pushing Gemini, but also open models with the Gemma 4 family. These are important for building an ecosystem of AI developers."

That ecosystem argument is the one worth watching. For anyone evaluating open models in production, the licensing question on Google's side just disappeared. What decides the race between Gemma 4, Qwen 3.5, and Meta's Llama won't be benchmark margins. It'll be tooling depth, fine-tuning community, and which model family gets you from prototype to production with the fewest surprises.

Alibaba Releases Third Closed-Source AI Model in Three Days, Signals Profit-First Strategy

Three proprietary AI models in three days. That's what Alibaba shipped this week, with Qwen3.6-Plus landing Thursday as the third, Bloomberg reported. This one handles agentic coding and multimodal re

The Implicator

GitHub Reverses Policy, Will Train AI on Copilot User Data Starting April 24

GitHub announced on March 25 that it will begin using interaction data from Copilot Free, Pro, and Pro+ subscribers to train its AI models, reversing a policy that previously excluded all plan tiers f

The Implicator

OpenAI Returns to Open-Source Arena With U.S.-Made Models Aimed at China’s Rise

💡 TL;DR - The 30 Seconds Version 🚀 OpenAI releases gpt-oss-120b and gpt-oss-20b models under Apache 2.0 license—first open-weight models since 2019's GPT-2. 💻 The 20b model runs on laptops wi

The Implicator

Harkaram Grewal

New Delhi

Maps the India–Germany–U.S. AI triangle from New Delhi. Background in cross-market operations and business development. Writes about supply chains, enterprise adoption, and talent—the unsexy forces that actually move global AI.