Google is giving its open AI strategy a serious upgrade with Gemma 4, a new family of models built for advanced reasoning and “agentic” workflows — the kind of AI that can not only answer questions but also call tools, run code, and orchestrate multi-step tasks on its own. And unlike many high-end models, Google is putting all of this under a permissive Apache 2.0 license, opening the door for startups, enterprises, and solo developers to use it commercially without jumping through legal hoops.

At a high level, Gemma 4 is meant to sit alongside Google’s proprietary Gemini line rather than replace it. Gemini remains the flagship service model you hit via APIs, while Gemma is the “run-it-yourself” sibling, with downloadable weights, local deployment, and full control over data and infrastructure. Google says Gemma has already been downloaded more than 400 million times, spawning over 100,000 community variants in what it calls the “Gemmaverse,” and Gemma 4 is clearly designed to push that ecosystem into a new phase.

The Gemma 4 lineup comes in four sizes: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture-of-Experts model, and a 31B dense model. On paper, those numbers might sound modest compared to the giant frontier models we hear about, but the story here is “intelligence per parameter.” Google is emphasizing that its 31B model is punching far above its weight: on Arena AI’s crowdsourced leaderboard, Gemma 4 31B debuts as the #3 open model in the world, while the 26B model takes the #6 spot, beating models up to roughly 20–30 times its parameter count. Arena’s team even calls Gemma 4 the top-ranked US open-source model, making it a serious contender for anyone who wants cutting-edge performance but doesn’t want to host a 300B-parameter behemoth.

Related /

• Gemma 4 under Apache 2.0 changes open AI forever

Where things get more interesting is what these models are actually built to do. Google describes Gemma 4 as “purpose-built for advanced reasoning and agentic workflows,” which is really shorthand for three capabilities that matter in practice: multi-step logic, tool use, and structured outputs. On the reasoning side, Gemma 4 shows big gains on benchmarks that stress math and instruction-following, which are the same ingredients you need to build reliable coding copilots, analysis agents, or decision support systems. For agentic workflows, the models come with native function calling, system instructions, and structured JSON output baked in, so a developer can wire Gemma 4 into an API, database, or automation engine and have it reliably choose tools, call them with the right arguments, and pipe the results back into the conversation.

Gemma 4 is also clearly aimed at developers who want local-first coding assistants. Google highlights “high-quality offline code generation,” essentially turning your workstation into a private AI pair programmer. The larger 26B and 31B models are optimized to fit on a single 80GB NVIDIA H100 in full bfloat16, but there are quantized versions tuned for consumer GPUs as well, which means serious coding agents and reasoning bots no longer have to live only in the cloud. The 26B model is a Mixture-of-Experts that only activates about 3.8B parameters per token, so it focuses on speed and tokens-per-second, while the 31B dense variant goes all-in on raw quality and is the better base if you plan to fine‑tune heavily for your own domain.

Multimodality is table stakes in 2026, and Gemma 4 doesn’t skip it. All four variants natively understand images and video, with support for variable resolutions and tasks like OCR, chart interpretation, and visual reasoning. On the smaller E2B and E4B models, Google goes a step further with native audio input for speech, which opens up use cases like offline voice assistants, call summarization, and speech-driven UI on phones and embedded devices. Out of the box, the models are trained on more than 140 languages, so developers in non‑English markets don’t have to fight the usual “fine-tune everything from scratch” battle to get acceptable results.

Another big focus is context length. The edge models (E2B and E4B) support up to a 128K token window, while the larger 26B and 31B options go up to 256K tokens. In practical terms, that means you can throw entire code repositories, long research papers, multi-chapter legal documents, or weeks of logs into a single prompt and still have room to reason about them. For things like RAG systems, doc review, or long‑horizon agents, that context budget is often more important than raw parameter size.

Where Gemma 4 really stands out compared to many rival open models is how aggressively Google has optimized it for real hardware instead of just benchmark charts. The E2B and E4B variants are engineered for phones, IoT boards, and low-power edge devices, activating only an “effective” 2B or 4B parameters at inference time to save RAM and battery while still delivering strong reasoning. Google co-designed these with the Pixel team and hardware partners like Qualcomm and MediaTek, and the models can run fully offline with near-zero latency on devices like Android phones, Raspberry Pi, and NVIDIA Jetson Orin Nano. In some configurations, E2B can fit in under about 1.5GB with aggressive quantization and still manage usable token speeds on hardware as modest as a Raspberry Pi 5.

That optimization work extends up the stack. Gemma 4 is already wired into Google’s own tooling: you can spin it up in Google AI Studio for quick experiments with the 31B and 26B models, or use AI Edge Gallery to explore mobile and embedded deployments with E2B and E4B. Android developers get integration through the AICore Developer Preview and Agent Mode in Android Studio, plus support via the ML Kit GenAI Prompt API to start baking local agents directly into apps. If you want to move beyond tinkering, Gemma 4 also slots into Vertex AI, Cloud Run, and GKE on Google Cloud, with options to scale out on TPUs, RTX-class GPUs, or sovereign cloud setups for regulated industries.

Outside Google’s own ecosystem, Gemma 4 lands with unusually broad day‑one support. You can download weights from Hugging Face, Kaggle, or Ollama, and run them through popular stacks like Transformers, TRL, vLLM, llama.cpp, MLX (for Apple Silicon), Ollama, LiteRT-LM, LM Studio, and more. NVIDIA is distributing Gemma 4 via its own NIM / RTX AI offerings, AMD support comes via ROCm, and there’s explicit tuning for everything from Blackwell data center GPUs down to Jetson edge devices. That spread makes it much easier for teams to fit Gemma 4 into existing pipelines rather than rebuilding infrastructure around a single vendor’s SDK.

Licensing is the other headline move. Earlier Gemma releases used a custom license that was technically open-weight but came with enough restrictions to create friction for some commercial use cases. With Gemma 4, Google is fully switching to the OSI-approved Apache 2.0 license, which is widely understood in the industry and gives developers broad freedom to modify, redistribute, and embed the models in commercial products without complex legal negotiations. Google frames this as a step toward “digital sovereignty,” arguing that organizations can now own their models end‑to‑end — deploy on‑prem, tune on proprietary data, and keep everything under their control while still benefiting from Google’s research.

There’s also a strong “responsible AI” angle. Google says Gemma 4 models go through the same infrastructure security and safety processes as its closed Gemini models, which is a nod to enterprises and governments that are wary of running open models in sensitive environments. The company points to earlier Gemma‑based collaborations, like INSAIT’s Bulgarian‑first BgGPT and Yale’s Cell2Sentence-Scale cancer research project, as examples of how open models can still be paired with serious oversight and domain‑specific guardrails. For Gemma 4 specifically, that likely means more thorough red‑teaming, safety filters, and documentation, though the details will matter as independent auditors dig in.

For developers, the launch also comes with the usual incentives and community hooks. There’s a “Gemma 4 Good” challenge on Kaggle to encourage projects that use the models for social impact, plus curated examples and model cards that document performance across a broad suite of benchmarks. The bigger picture is that Google wants Gemma 4 to be the default choice when someone asks, “Which open model should I start with for agents, reasoning, and on‑device AI?” — much like how LLaMA once became the default baseline for open large language models.

If you zoom out, Gemma 4 looks like a statement of intent. On one side, proprietary giants like Gemini, OpenAI’s latest GPTs, and other frontier systems are racing ahead with capabilities that are hard to match without huge compute budgets. On the other, there’s a rapidly maturing open ecosystem where performance gaps are closing, model sizes are shrinking, and the real differentiators are things like latency, cost, hardware coverage, and licensing. Gemma 4 is Google’s attempt to straddle that line: take pieces of its frontier research, compress them into efficient open models, and let the community run wild with them — from offline agents on your phone to serious reasoning engines running on a single GPU.