Gemma 4 Models Released as Open Weight AI
Google released its Gemma 4 models today. It’s a new family of open weight multimodal AI systems, and they arrive under an Apache 2.0 license. The announcement came in a Google AI blog post co-authored by DeepMind VP of Research Clement Farabet and Director of Product Management Olivier Lacombe. They called the release the most capable open models Google has ever shipped, built for advanced reasoning and agentic workflows.
Since the first Gemma generation launched, developers have downloaded the models over 400 million times and built more than 100,000 community variants. Google calls that collection the Gemmaverse. The new family carries that momentum forward with a full open source license that drops the restrictions of earlier versions. Google’s Open Source blog said the Apache 2.0 terms give developers autonomy, control, and clear rights to modify and redistribute the weights.
The whole Gemma 4 line shares the same research foundation as Google’s proprietary Gemini 3 models. That means the open models inherit the same safety infrastructure, with a direct path from lab to community. Every Gemma 4 model handles text and image input. The smallest variants also add native audio understanding, so they can process speech and environmental sound. Video and variable resolution image processing work across all sizes, and the context window runs to 128K tokens on edge models and 256K on the larger ones.
Five Model Sizes From Mobile To Server
Google built this family around efficiency per byte, not just raw scale. The family includes five distinct sizes, covering everything from a phone to a developer workstation. The two smallest, E2B and E4B, use Per-Layer Embeddings to keep the effective active parameter count low while still handling multimodal perception. E2B runs as a 2.3 billion effective parameter model despite holding 5.1 billion total parameters. E4B works as a 4.5 billion effective model inside an 8 billion parameter envelope.
The 12B Unified model drops dedicated encoders entirely. It fuses text, image, audio and video processing into a single transformer stack, so the whole thing is a 12 billion parameter package that fits on a consumer GPU. Next up is the 26 billion parameter mixture of experts model, which activates only 4 billion parameters per forward pass to save compute and memory. The largest 31B dense model keeps the full parameter count live, which gives it the edge in code generation and complex reasoning.
Every model uses a hybrid attention design, mixing local sliding window attention with full global attention at key layers. The official model card explains how this keeps the memory footprint of a lightweight model while still preserving the long range awareness a 256K token context needs. The edge models target phones, Raspberry Pi, and Jetson Nano boards. The larger models are tuned for workstation GPUs, so a student can turn a laptop into a private AI lab.
Gemma 4 Benchmarks Show Top Tier Efficiency
Benchmark results put the new models among the most efficient open AI releases out there. On the Arena AI text leaderboard as of early April 2026, the 31B dense model scored 1452 points and sits as the third highest open model in the world. The 26B mixture of experts model hit 1441 points and ranks sixth, beating models twenty times its size. Those figures come straight from the model card Google DeepMind published.
Reasoning scores jumped a lot over the previous Gemma generation. The 31B variant reached 89.2 percent on the AIME 2026 mathematics benchmark without tools, and the 26B MoE scored 88.3 percent. For comparison, Gemma 3 27B only managed 20.8 percent. On the GPQA Diamond science knowledge test, the same large models passed 84.3 percent and 82.3 percent. Code generation on LiveCodeBench v6 hit 80 percent for the 31B and 77.1 percent for the 26B.
Agentic benchmarks point to a practical use case. The 31B and 26B models scored above 85 percent on the τ2-bench retail task, which tests function calling and tool use. Even the tiniest E4B model managed 57.5 percent on that agentic test, while the prior generation’s 27B model reached only 6.6 percent. Those numbers suggest Gemma 4 models can power reliable autonomous agents that call APIs, generate structured JSON, and follow system prompts without needing the cloud.
Gemma 4 Models Available Now Under Apache 2 0
The models are available for download right now on Hugging Face and through Google AI Studio. Developers can fine-tune them with familiar frameworks and run inference entirely offline. The Apache 2.0 license covers all sizes, both pre-trained and instruction-tuned variants, which gives the community a clear open source foundation for commercial and research work alike. The launch lands the same week as Moonshot AI’s Kimi K3 open-weight release, another heavyweight open model vying for developer attention, though Kimi K3 ships under a custom license rather than Apache 2.0.
The release also marks the first time a Gemma generation has shipped under an OSI-approved license. Google’s open source leaders called it a direct response to what the community had been asking for. With those tools in hand, the Gemmaverse can grow from 100,000 variants into a much wider ecosystem of offline agents, sovereign AI services, and local-first coding assistants.
Frequently Asked Questions About Gemma 4
What sizes do the Gemma 4 models come in?
The family has five sizes. E2B runs as an effective 2.3 billion parameter model, tuned for mobile devices and single board computers. E4B scales that up to 4.5 billion effective parameters. The 12B Unified model processes text, images, audio and video inside a single encoder-free architecture. A 26 billion parameter mixture of experts model activates 4 billion parameters per token to save compute. The largest 31 billion parameter dense model serves workstations and servers.
How does the Apache 2.0 license help developers?
The license gives developers full rights to use, modify and distribute the models without restrictive terms attached. Google pointed to three gains: autonomy to build derivative models, control to run everything on private hardware without cloud dependency, and legal clarity that comes from using industry standard license text. Previous Gemma generations carried custom terms. This shift lets enterprises and public institutions deploy the models with the same confidence they’d have in other Apache 2.0 projects like Kubernetes and Android.
Can I run Gemma 4 models on my phone?
Yes. The E2B and E4B sizes are built specifically for on-device execution. They include native audio input for voice interaction, support variable resolution vision, and run with near-zero latency on Android phones. Google’s AI Edge tooling makes it possible to test them offline directly from a browser. The 12B and larger models need a laptop GPU or a cloud instance, but they still fit within consumer hardware budgets.
