Google dropped Gemma 4 on April 2, 2026, A full generational jump in what open models can do at their parameter range and the first time in the Gemma family's history that one ships under Apache 2.0, meaning commercial use without permission-seeking.
Some context: since Gemma's first generation, developers have downloaded the models over 400 million times and built more than 100,000 variants.
Four Models, One Family
Gemma 4 is a family of four, each aimed at a different point in the hardware spectrum.
E2B : Effective 2 billion active parameters. Runs on smartphones, Raspberry Pi, Jetson Orin Nano. 128K context window. Handles images, video, and audio. Built for battery and memory efficiency.
E4B : Effective 4 billion active parameters. Same hardware targets, higher reasoning quality. About 3x slower than E2B, but noticeably more capable. Also supports images, video, and audio. Up to 4x faster than prior versions, 60% less battery.
26B MoE : 26 billion total parameters, but only 3.8 billion activate during inference. Context window up to 256K tokens. Ranked 6th among all open models on the Arena AI Text Leaderboard. Quantized versions run on consumer GPUs.
31B Dense : The flagship. Full dense architecture. 256K context. Currently ranked 3rd among open models on Arena AI. Fits unquantized on a single 80 GB H100; quantized versions run on consumer hardware. The obvious fine-tuning base.
One thing to notice: E2B and E4B both handle audio input natively. The 26B and 31B do not. If speech recognition is part of your application, the edge models are your only option in this family.
Benchmarks: The Numbers That Matter
Google says Gemma 4 outperforms models 20 times its size. Sounds like marketing. Third-party data from Artificial Analysis makes it harder to wave away.
On GPQA Diamond (scientific reasoning), the 31B scores 85.7% in reasoning mode. Second-best among open models under 40 billion parameters, just behind Qwen3.5 27B at 85.8%. The efficiency angle is also worth noting: the 31B generates roughly 1.2 million output tokens in that evaluation, versus 1.5 million for Qwen3.5 27B. Less compute for roughly equal quality.
The 26B MoE scores 79.2% on GPQA Diamond, putting it ahead of OpenAI's gpt-oss-120B at 76.2%. That is a 94-billion-parameter gap between those two models.
The agentic tool use numbers are where it gets genuinely interesting. On τ2-bench (Retail), the 31B scores 86.4% and the 26B scores 85.5%. Gemma 3 27B scored 6.6% on the same benchmark. Whatever changed in how these models handle multi-step tool use, it was not incremental.
Math and coding follow a similar pattern. On AIME 2026, the 31B and 26B reach 89.2% and 88.3%, versus Gemma 3 27B's 20.8%. On LiveCodeBench v6, the 31B scores 80.0% and the 26B scores 77.1%. Gemma 3 27B was at 29.1%.
The edge models are more modest. E4B hits 52.0% on LiveCodeBench and 58.6% on GPQA Diamond. Reasonable for a model designed to fit on a phone.
Architecture: What Actually Changed
Gemma 4 comes from the same research stack as Gemini 3, Google's closed model family. The reasoning and math benchmark jumps suggest the knowledge transfer from that training actually worked, not just a talking point.
The MoE design in the 26B model is worth understanding. The total parameter count is 26 billion, but only 3.8 billion activate on any given forward pass. In practice, this gets you near-31B quality at a fraction of the inference cost. Token generation should be faster than the dense model, with some trade-off in raw quality.
Both the 26B and 31B support function calling, structured JSON output, and system instructions natively. Gemma 3 was awkward for agentic use. Gemma 4 was built for it from day one.
The models also cover over 140 languages, which opens up a wider range of deployment contexts without additional localization work.
On-Device Deployment: What "Completely Offline" Actually Means
The E2B and E4B models run fully offline on Android, Raspberry Pi, and NVIDIA Jetson Orin Nano. Google worked with Qualcomm Technologies and MediaTek on hardware optimization. Android developers can start prototyping agentic flows through the AICore Developer Preview now, and code written for Gemma 4 will be forward-compatible with Gemini Nano 4 devices coming later this year.
Why does running offline matter? Three things: latency (no network round trip), privacy (data stays on device), and reliability (no API dependency). For healthcare apps, legal review tools, or anything touching sensitive user data, local inference is not just a nice-to-have.
One caveat: the AICore Developer Preview is still a preview. Tool calling, structured output, and thinking mode through the Prompt API are coming during the preview period, not available at launch. If you are building for production Android deployment today, check what is actually ready before you commit to an architecture.
Where to Get It and What Runs It
Gemma 4 is on Hugging Face, Kaggle, and Ollama now. Google AI Studio has the 31B and 26B. Google AI Edge Gallery covers the E4B and E2B.
Framework support at launch is broad: Hugging Face Transformers, TRL, Transformers.js, Candle, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM and NeMo, LM Studio, Unsloth, SGLang, Keras, and more. You can train on Google Colab, Vertex AI, or a consumer gaming GPU.
For production, the options are Vertex AI, Cloud Run, GKE, or TPU-accelerated serving. NVIDIA hardware from Jetson Orin Nano to Blackwell GPUs is supported. AMD GPUs work via ROCm.
The Apache 2.0 License: Why It Actually Matters
Every previous Gemma release shipped under a Google proprietary license. Gemma 4 is the first to go Apache 2.0.
What that means concretely: you can build and sell products with Gemma 4, modify it, fine-tune it, redistribute it, and deploy it, all without negotiating Google's terms. For startups and solo developers, that is one less legal headache when taking something to market.
What to Use It For (And What to Watch)
The jump in agentic tool use is the most significant thing about this release. If you are building multi-step reasoning pipelines, function calling workflows, or autonomous agents, Gemma 4 is a different category of model than Gemma 3 was.
The 31B dense model, with its strong scores across reasoning, coding, and science benchmarks, is a solid starting point for fine-tuning on domain-specific data. Familiar tooling (Hugging Face, Unsloth, Colab, Vertex AI) means the fine-tuning workflow will not surprise you.
The honest caveat: benchmarks are controlled. τ2-bench numbers are encouraging, but whether the agentic improvements hold up when tool schemas are messy and information is partial - that is something the community will figure out over the next few weeks. Worth watching.
Getting Started
# Pull with Ollama
ollama pull gemma4:31b
ollama pull gemma4:26b-moe
# Or grab from Hugging Face
pip install transformers
Google AI Studio has the 31B and 26B available directly in browser - no local setup needed to start experimenting.
For Android, the AICore Developer Preview requires opt-in, after which you can trigger model download directly to a supported test device.
There is also a Gemma 4 Good Challenge on Kaggle if you learn better by building toward a concrete problem.
Top comments (0)