Google Just Made Its Strongest Open Model Truly Free
Google released Gemma 4 on April 2, 2026 — and the release did something no previous Gemma version had done: it shipped under a fully permissive Apache 2.0 license. Previous Gemma generations were open-weight but carried Google-specific commercial restrictions. Gemma 4 removes every one of them. No usage caps, no MAU limits, no acceptable-use policies beyond the standard open-source terms. For developers and enterprises building on open models, that licensing change alone would have been significant news. The benchmark performance on top of it is what turned Gemma 4 into the most-discussed open model release of 2026.
The numbers are hard to ignore: the Gemma 4 31B Dense model reached third place on Arena's global text leaderboard — outcompeting models with parameter counts 20 times larger. The community has already downloaded Gemma models over 400 million times across all generations, building more than 100,000 variants in what Google calls the Gemmaverse. Gemma 4's first-week download velocity reportedly exceeded every prior Gemma release.
What is Gemma 4? Gemma 4 is Google DeepMind's fourth-generation open model family, released April 2, 2026. Built on the same research foundation as the proprietary Gemini 3, it comes in four sizes — from a 2B edge model to a 31B workstation model — with multimodal input (text, image, video, and audio), context windows up to 256K tokens, and an Apache 2.0 license with no commercial restrictions.
The Four Gemma 4 Models Explained
Gemma 4 is not a single model — it is a coordinated family designed to cover every deployment scenario, from an offline smartphone to a cloud server. Each variant has a specific purpose.
Gemma 4 E2B — Edge and On-Device
The E2B model carries approximately 2.3 billion effective parameters (total including embeddings reaches about 5.1B). It is a multimodal model with native audio input, optimized for running completely offline on phones, Raspberry Pi boards, and NVIDIA Jetson Orin Nano. Context window: 128K tokens. Google built it in close collaboration with Qualcomm Technologies and MediaTek, and Android developers can prototype agentic flows using it today through the AICore Developer Preview. If your use case requires zero-latency, zero-connectivity inference on a device a user holds in their hand, this is the Gemma 4 variant to test first.
Gemma 4 E4B — The Developer Sweet Spot
The E4B model sits at approximately 4 billion effective parameters. It runs on integrated graphics or any GPU with 8 GB or more of VRAM — hardware most developers already own. It beats Gemma 3's 27B model across benchmarks despite having a fraction of the parameter count, with AIME 2026 math scores of 69.4%. Native audio input is included. Context window: 128K tokens. For the majority of developers who want local inference without investing in server hardware, the E4B is the primary reason to evaluate Gemma 4 over alternatives.
Gemma 4 26B A4B — MoE Efficiency at Scale
The 26B model uses a Mixture-of-Experts (MoE) architecture that activates only 3.8 billion parameters per inference token despite having 26 billion total parameters. In practice, this means near-31B quality output at E4B-class VRAM cost during inference — quantized versions run on consumer GPUs. Context window: 256K tokens. It is available as a fully managed, serverless deployment on Google Cloud's Model Garden. For production teams who need frontier-quality reasoning without paying for 30B+ dense model inference costs, this is the most cost-effective entry point in the Gemma 4 family.
Gemma 4 31B Dense — Flagship Reasoning
The 31B Dense model is the top of the Gemma 4 lineup and the variant behind the headline benchmark numbers. It reached third place on Arena's text leaderboard globally. Key scores: 89.2% on AIME 2026 (advanced mathematics), 84.3% on GPQA Diamond (graduate-level science reasoning), and 80.0% on LiveCodeBench v6 (coding). Context window: 256K tokens. It is deployable on NVIDIA Blackwell GPUs via Cloud Run — which provides 96 GB of vGPU memory — and fine-tunable through Vertex AI Training Clusters with optimized SFT recipes. This is the model to benchmark if you are evaluating whether an open model can replace a proprietary API in your production stack.
Benchmark Comparison: Gemma 4 vs Llama 4 vs Mistral
| Model | Params (Active) | AIME 2026 | GPQA Diamond | LiveCodeBench v6 | License | Context |
|---|---|---|---|---|---|---|
| Gemma 4 31B Dense | 31B | 89.2% | 84.3% | 80.0% | Apache 2.0 | 256K |
| Gemma 4 26B MoE | 3.8B active | ~85% | ~81% | ~76% | Apache 2.0 | 256K |
| Gemma 4 E4B | 4B | 69.4% | — | — | Apache 2.0 | 128K |
| Llama 4 Scout | 17B active / 109B total | — | 74.3% | — | Meta Custom* | 10M |
| Llama 4 Maverick | 400B total MoE | — | — | — | Meta Custom* | 1M |
| Mistral Small 4 | 6B active / 119B total | — | — | Leads on efficiency | Apache 2.0 | 256K |
*Llama 4 uses Meta's custom license, which restricts commercial deployment for applications exceeding 700 million monthly active users. Gemma 4's Apache 2.0 has no such restrictions.
Where Gemma 4 Wins — and Where It Does Not
Clear Strengths
Reasoning and mathematics: The 31B Dense model's 89.2% on AIME 2026 is the headline number. For agentic workflows that require multi-step logical reasoning — planning, tool calling, structured output — Gemma 4 is among the strongest open options available as of April 2026.
On-device deployment: No other frontier-adjacent open model family runs on a Raspberry Pi and a workstation server under the same license. The E2B model is the only practical choice for fully offline, zero-latency on-device AI among models at this quality tier.
Licensing freedom: Apache 2.0 with no MAU restrictions is the cleanest commercial license in the open-model space. If you are building a product that could scale beyond 700 million users, Gemma 4 and Qwen 3.5 are the only frontier-adjacent families without license complications.
Multimodality: All four Gemma 4 models process images and video. The E2B and E4B additionally support native audio input. The 26B and 31B support over 140 languages. For global products that handle mixed-media inputs, this breadth is unusual at open-model price points.
Ecosystem integration: Gemma 4 is available immediately on Hugging Face (transformers, llama.cpp, MLX, WebGPU, Rust), Ollama, Kaggle, Vertex AI, Cloud Run, and GKE. Google's Agent Development Kit (ADK) provides native support for building agentic systems on top of Gemma 4 with function calling and structured output.
Real Limitations
Context window vs Llama 4: The Gemma 4 31B's 256K context window is impressive for an open model, but Llama 4 Scout ships with a 10 million token context window — the largest of any open model. For use cases that require whole-codebase reasoning across massive repositories or very long document analysis, Llama 4 Scout has no peer among open models.
Multimodal depth vs Llama 4: On multimodal benchmarks specifically, Llama 4 holds a narrow edge. If image and video understanding is your primary workload rather than a secondary capability, Llama 4 warrants a separate evaluation.
Coding specialists: On coding-specific benchmarks, GLM-5.1 and the anticipated DeepSeek V4 lead the field. Gemma 4's coding performance is strong at 80.0% on LiveCodeBench v6, but it is not the top coding model in the open-source space.
How to Access and Deploy Gemma 4
Gemma 4 is available across multiple platforms with no account requirements for the base models. The most common deployment paths are listed below.
- Hugging Face: All four models are listed under
google/gemma-4-*. The any-to-any pipeline in Transformers supports the E2B and E4B models with a singlepipeline("any-to-any", model="google/gemma-4-e2b-it")call. - Ollama: Local inference on macOS, Linux, and Windows without writing any code. Pull and run the E4B or 26B MoE model directly.
- Kaggle: Free GPU notebooks for experimentation without local hardware requirements.
- Google Cloud Vertex AI: Deploy to your own Vertex AI endpoints with fine-tuning support via NeMo Megatron. The 26B MoE is available as serverless on Model Garden.
- Cloud Run: Serverless GPU deployment with NVIDIA RTX PRO 6000 Blackwell GPUs, 96 GB vGPU memory, and scale-to-zero billing.
- Google Kubernetes Engine (GKE): For teams that need custom autoscaling, security controls, and microservices integration.
Explore available open-source AI models on SwitchTools to compare Gemma 4 against other models you can self-host today.
Who Gemma 4 Is NOT For
Gemma 4 is not the right choice if your primary workload requires context windows beyond 256K tokens — Llama 4 Scout's 10 million token window is in a different category for that specific need. It is also not the top pick for pure coding benchmarks, where GLM-5.1 currently leads. If you are building a non-technical product where raw API cost per call is the only variable that matters and you have no data privacy requirements, a managed API from Anthropic or OpenAI may be simpler to operate than self-hosting even the E4B. Finally, teams without any ML infrastructure experience may find Google Cloud's managed options easier than bare self-hosting via Ollama or Hugging Face, even though both are straightforward for developers comfortable with terminals.
Frequently Asked Questions
Is Gemma 4 truly open source in 2026?
Yes. Gemma 4 is the first Gemma generation released under the OSI-approved Apache 2.0 license. Previous versions used Google's custom "Gemma Open" license, which allowed open weights but imposed commercial restrictions. Apache 2.0 allows full modification, redistribution, and commercial deployment with no usage-based limitations.
How does Gemma 4 compare to Llama 4 for local deployment?
Gemma 4 is significantly more accessible for local use. The E4B model runs on any GPU with 8 GB of VRAM. Llama 4 Scout requires a minimum of 24 GB VRAM even with aggressive quantization, making it a server-only option for most developers. On reasoning and math benchmarks, Gemma 4 31B leads Llama 4 Scout. On multimodal tasks and raw context length, Llama 4 has the advantage.
What hardware do I need to run Gemma 4 locally?
The E2B model runs on smartphones, Raspberry Pi, and Jetson Orin Nano. The E4B runs on any GPU with 8 GB or more of VRAM — including integrated laptop graphics. The 26B MoE model, despite its full parameter count, activates only 3.8B parameters during inference, making it runnable on consumer GPUs with quantization. The 31B Dense model requires a proper workstation GPU or cloud deployment.
Can I use Gemma 4 for commercial products without restrictions?
Yes. The Apache 2.0 license imposes no restrictions on commercial use, distribution, or modification beyond standard attribution requirements. Unlike Llama 4's custom Meta license — which restricts deployment for services with more than 700 million monthly active users — Gemma 4 places no such ceiling on scale. This makes it the strongest license in the frontier open-model space as of April 2026.
Which Gemma 4 model should I start with?
For most developers, the E4B is the starting point — it runs on hardware you already own, supports multimodal input, and delivers benchmark performance that beats the previous generation's 27B model. If you need production-grade reasoning quality and have server or cloud access, evaluate the 26B MoE first for its compute efficiency before committing to the full 31B Dense. Explore the full AI models directory on SwitchTools to see how Gemma 4 fits alongside other open and closed models available today.