Google Gemma 4: The Open-Source AI Model That Enterprises Can Finally Trust
Apache 2.0 licensing removes the legal barriers that kept businesses away — and the benchmark numbers suggest the wait was worth it
For years, the phrase "open AI" has come with an asterisk. Models released under custom proprietary licenses — with terms their creators could rewrite at will — offered the appearance of openness while quietly reserving the right to pull the rug out from under developers. On April 2, 2026, Google tore up that asterisk.
With the release of Gemma 4, Google has delivered its most capable open-weight model family to date — debuting at #3 among all open models globally on the Arena AI Text Leaderboard — under the Apache 2.0 license for the first time in the Gemma series' history. That licensing shift, easy to overlook beside the benchmark numbers, may prove to be the more consequential announcement. It transforms Gemma from a model developers admired but hesitated to build on into one they can commit to without legal reservation.
A Family Built for Every Device, From Raspberry Pi to H100
Gemma 4 arrives as a family of four models, each targeting a different tier of the hardware spectrum. The Effective 2B (E2B) and Effective 4B (E4B) models are engineered for edge deployment — smartphones, Raspberry Pi devices, and Jetson Orin Nano hardware — with a 128,000-token context window and support for audio input and speech recognition. The larger 26B Mixture-of-Experts (MoE) and 31B Dense models step up to a 256,000-token context window and target workstations and server-grade infrastructure, with the 31B Dense fitting unquantized on a single 80GB NVIDIA H100 GPU.
All four models share a common foundation: they are built on the same underlying technology as Google's proprietary Gemini 3. Every variant supports offline operation, vision inputs including images and video, multimodal reasoning, function calling, structured JSON output, system instructions for agentic workflows, and more than 140 languages. Google DeepMind describes the larger models as offering "unprecedented intelligence-per-parameter" and "frontier intelligence on personal computers" — claims that, unusually for AI marketing, are largely corroborated by independent data.
The benchmark numbers are striking. On AIME 2026 mathematics, the 31B Dense model scores 89.2% and the 26B MoE scores 88.3% — compared to just 20.8% for Gemma 3 27B. On LiveCodeBench v6 coding tasks, the 31B reaches 80.0% and the 26B hits 77.1%, against Gemma 3 27B's 29.1%. On GPQA Diamond scientific knowledge, independent benchmarks from Artificial Analysis place the 31B at 85.7% in reasoning mode — second only to Qwen3.5 27B at 85.8% among open models under 40 billion parameters. Google claims Gemma 4 "outperforms models 20 times its size," and the Arena AI Elo scores — 1,452 for the 31B and 1,441 for the 26B MoE, versus 1,365 for Gemma 3 27B — suggest the leap is real.
The License That Changes Everything
Performance improvements, however impressive, are table stakes in today's AI landscape. What makes Gemma 4 genuinely different is what Google changed in the fine print.
Previous Gemma versions shipped under a custom Google proprietary license that developers widely found problematic. As Ars Technica reported, "The Gemma 3 license had a strict prohibited-use policy that Google could update unilaterally, and it required developers to enforce Google's rules across all Gemma-based projects. It could even be read to transfer the license to other AI models created with synthetic data produced by Gemma. This made many devs apprehensive about building with Google's open models."
Sources & References
More from Pulse AI
Cursor 3 Launches an AI Agent Platform to Take On Claude Code and OpenAI's Codex
Cursor has launched Cursor 3, a new agent-first coding platform developed under the code name 'Glass,' putting the startup in direct competition with Anthropic's Claude Code and OpenAI's Codex. The new interface — built from scratch and integrated alongside the existing Cursor IDE — lets developers spin up autonomous AI agents from desktop, mobile, web, Slack, GitHub, and Linear, with all agents surfaced in a unified sidebar. Cursor also introduced Composer 2, its own frontier coding model, signaling a push for independence from third-party AI providers. The launch comes as Cursor faces mounting pricing pressure from better-funded rivals and is reportedly raising capital at a $50 billion valuation.