Google’s new Gemma 4 family is the company’s most capable open AI model to date, combining frontier‑level reasoning and multimodal skills with an unusually permissive Apache 2.0 license that lets developers download, modify, and deploy it almost anywhere. Built on the same research foundation as Google’s proprietary Gemini 3 models, Gemma 4 ranges from tiny variants that run fully offline on phones to 31‑billion‑parameter systems that already rank among the top open models on community leaderboards.
What Gemma 4 is, and how it fits with Gemini
Gemma is Google’s line of “small, open” models; Gemini is the company’s large, proprietary, subscription‑backed AI family that powers products like Search, Workspace and Google Cloud services. Gemma 4 is the latest generation of those open models, built on the same research and architecture foundations as Gemini 3 but released for developers to run on their own infrastructure.
Google’s launch blog describes Gemma 4 as “our most intelligent open models to date,” designed specifically for advanced reasoning and “agentic workflows”, AI agents that can plan multi‑step tasks, call tools and interact with APIs. While Gemini sits behind Google’s cloud APIs and consumer products, Gemma 4 can be downloaded as model weights, fine‑tuned, and embedded in third‑party apps, edge devices or private data centers.
That distinction matters for enterprises and researchers who want “digital sovereignty”: control over where models run and how data is handled, without sending everything back to Google’s servers.
Model lineup: from “effective” 2B to 31B
Gemma 4 launches as a four‑model family tailored to different hardware and workloads.
Google and partner write‑ups describe the lineup as:
- E2B (“Effective 2B”): an ultra‑efficient edge model that activates roughly 2 billion parameters at inference time, designed for phones, IoT devices, browsers, and single‑GPU laptops.
- E4B (“Effective 4B”): a slightly larger edge model with more capacity but still tuned for mobile and embedded deployments.
- 26B MoE (Mixture of Experts, sometimes branded “A4B”): a sparsely activated model with 26 billion parameters overall but around 4B active per token, aimed at high‑throughput, advanced reasoning with good efficiency.
- 31B Dense: a full 31‑billion‑parameter dense model designed to “bridge the gap between server‑grade performance and local execution,” running on workstation‑class GPUs and accelerators.
Google says the 31B dense variant currently ranks as the #3 open model on the Arena AI community text leaderboard, with the 26B MoE at #6, both outperforming much larger open models on standard benchmarks. For developers, that “intelligence‑per‑parameter” means strong performance without needing H100 clusters.
Multimodal, long‑context and agent‑ready
Unlike early Gemma releases that were text‑only, Gemma 4 is fully multimodal across the lineup.
Google’s model card and partner blogs highlight that:
- All Gemma 4 models accept text and images (and in some setups, video frames), generating text outputs.
- The smaller E2B and E4B variants additionally support audio input for on‑device speech recognition and understanding.
- The vision stack uses improved encoders with variable aspect ratios and configurable image token counts, optimizing the trade‑off between speed, memory, and quality.
- Gemma 4 models can handle long context windows, 128,000 tokens for the small edge models and 256,000 tokens for the larger 26B and 31B tiers, enabling long documents, chat histories or multi‑file codebases.
On the “agentic” side, the model overview notes built‑in function‑calling, improved multi‑step reasoning and “native system prompt support” so developers can structure conversations and tool‑use more reliably. Google positions Gemma 4 as a foundation for autonomous agents that can read docs, call APIs, orchestrate workflows and run on‑device when needed.
Coding is a major focus: Google and Hugging Face both report significant gains on code benchmarks, with first‑party fine‑tuned variants aimed at code completion, debugging and agent‑based code refactoring.
Mobile‑first AI: running Gemma 4 on phones and PCs
One of the headline goals for Gemma 4 is to make serious AI run well on everyday devices.
Google says E2B and E4B were “engineered from the ground up” for compute and memory efficiency, activating only their effective parameter footprint during inference to conserve RAM and battery. In collaboration with the Pixel team and chipmakers like Qualcomm and MediaTek, these edge models can run completely offline with “near‑zero latency” on phones, Raspberry Pi boards and Jetson‑class embedded devices.
For developers, that means:
- Android support via AICore developer previews, with a path to future Gemini Nano 4 integration.
- Browser‑side inference for smaller tasks, with Gemma 4 tuned for WebGPU and Chrome‑based deployments.
- Easy deployment on a single GPU or TPU in workstations for the 26B and 31B models, as SourceForge and Google’s docs emphasize.
That mobile‑first design is part of a wider industry push to move more AI “to the edge,” reducing latency, bandwidth costs and privacy risk by keeping inference local when possible.
The Apache 2.0 license: why it matters
Perhaps the most consequential choice is not technical but legal. Gemma 4 is released under the standard Apache 2.0 open‑source license, a shift from previous Gemma models that used a custom “Gemma license.”
Engadget and Ars Technica note that Apache 2.0 gives developers broad freedom: you can use Gemma 4 commercially, modify it, fine‑tune it on your own data, and redistribute derivatives without special revenue‑sharing terms or use‑case carveouts. Google explicitly pitches this as a foundation for “complete developer flexibility and digital sovereignty,” saying it lets customers “deploy securely across any environment, whether on‑premises or in the cloud.”
In a competitive landscape where many “open” models have restrictive community licenses, Apache 2.0 puts Gemma 4 closer to Meta’s early LLaMA 2/3 style of openness, but with stronger multimodal and agent capabilities out of the box.
What developers can do with Gemma 4, and the trade‑offs
Between the small and large variants, core use cases include:
- On‑device assistants: personal chatbots, transcription tools, and multimodal note‑takers running entirely on phones, laptops, or kiosks, using E2B/E4B for privacy‑sensitive workflows.
- Enterprise copilots and agents: 26B/31B models fine‑tuned on internal documents to answer questions, draft reports, route tickets or orchestrate multi‑step tasks across business systems.
- AI coding tools: IDE integrations where Gemma 4 suggests code, writes tests, or powers “vibe coding” workflows in which developers describe intent and the model generates scaffolding.
- Vision and data extraction: OCR, chart understanding, document triage and basic video understanding, leveraging Gemma 4’s improved vision stack and long context for multi‑page inputs.
The trade‑offs are typical of open models but worth emphasizing. While Gemma 4 scores well on public leaderboards, Google’s model card and Hugging Face’s analysis note that it can still hallucinate, carry training data biases, and misinterpret edge‑case inputs. Safety filters and guardrails are weaker than in tightly managed hosted services like Gemini Advanced and deploying Gemma 4 in production requires your own security, monitoring and compliance frameworks.
For organizations already invested in Google Cloud, Gemma 4 can be run via hosted endpoints; for those prioritizing sovereignty, its open weights and Apache license mean they can keep everything inside their own VPCs or even offline.
For now, Gemma 4’s message to the AI world is clear: Google wants to compete not just in closed, subscription AI, but in the open model arena too, and it’s willing to ship serious capabilities in a package developers can actually own, shape and run themselves.
