On April 2, 2026, Google DeepMind released Gemma 4, the first major revision of its open model family in a year. The most consequential change is not capability but licensing: the entire lineup ships under Apache 2.0, dropping the custom Gemma terms used for the first two generations. For the first time, a Google open model falls inside a standard open-source license — no extra usage restrictions, no vendor terms that can change unilaterally, and nothing novel for a corporate legal team to review.
The lineup itself covers the full deployment spectrum in one release: E2B and E4B edge models, a 26B mixture-of-experts model that activates only about 3.8B parameters at inference, and a 31B dense model — all built on the same research and technology behind Gemini 3, natively covering more than 140 languages, with up to 256K of context. Google’s positioning fits in one line: its most capable open models to date, purpose-built for advanced reasoning and agentic workflows.
From Custom License to Apache 2.0
Previous Gemma generations were free for commercial use, but under Google’s custom license: users had to accept attached terms, with constraints on naming and use. Individual developers barely noticed; enterprise platform teams did. A custom license means legal review, vendor risk assessment, and the standing possibility that the terms change under you. Ars Technica framed the Apache 2.0 switch as the core news of the release, not marketing garnish.
The timing matters too. An opening-of-year outlook we published in January was still tracking reports that Meta might take its Avocado frontier model closed-source. Three months later, Google answered in the opposite direction with a plain Apache 2.0 grant. The open-weights camp is splitting, and Gemma 4 has planted its flag squarely on the fully open side.
Four Sizes: From Phones to a Single H100
Gemma 4’s design logic is one research foundation, several deployment targets:
- E2B (~2B effective parameters) and E4B (~4.5B effective): built for phones, Raspberry Pi, and NVIDIA Jetson Orin Nano, with near-zero-latency offline execution, co-engineered with the Google Pixel team, Qualcomm, and MediaTek
- 26B MoE: latency-oriented, 26B total parameters but roughly 3.8B active at inference, balancing quality against cost
- 31B dense: quality-oriented and positioned by Google as a strong base for fine-tuning
On hardware: the bf16, unquantized versions of the two larger models fit on a single 80GB H100, and quantized builds run on consumer GPUs. Context windows run 128K on the edge models and up to 256K on the larger ones.
Benchmark Numbers and Arena Rankings
Per Hugging Face’s day-one evaluation of the instruction-tuned models: the 31B scores 85.2% on MMLU Pro, 89.2% on AIME 2026, 84.3% on GPQA Diamond, 80.0% on LiveCodeBench, and a Codeforces ELO of 2150. The 26B MoE lands at 82.6% MMLU Pro, 88.3% AIME 2026, 77.1% LiveCodeBench, and 82.3% GPQA Diamond. Most numbers sit clearly above the previous Gemma 3 27B.
The rankings are the more intuitive read: as of April 1, the 31B sat at number three among open models on the Arena text leaderboard and the 26B at number six, with the MoE variant posting an estimated LMArena score around 1441 on roughly 4B active parameters. Google’s claim that it “outcompetes models 20x its size” is marketing — but an open model reaching the open-source top tier on edge-class active parameters is real economics.
Agentic Workflows and Multimodal I/O
On capability: every variant natively processes image and video input at variable resolutions, with OCR and chart understanding; E2B and E4B add native audio input. For agentic work, the models support GUI element detection, object detection emitted as native JSON bounding boxes, multimodal function calling, and a thinking mode — effectively building the perception layer for “read the screen, click the button, call the tool” directly into the weights.
The ecosystem showed up on day one: weights are downloadable from Hugging Face, Kaggle, and Ollama, with first-day support in transformers, llama.cpp, MLX, Mistral.rs, TRL, Unsloth, and Vertex AI. On Android, Gemma 4 powers Agent Mode in Android Studio and is forward-compatible with a future Gemini Nano 4 through AICore. An experimental DiffusionGemma (26B base, 3.8B active) generates text via discrete diffusion, exceeding 1,100 tokens per second on an H100 in FP8, though it still trails the autoregressive model on most benchmarks (77.6% vs 82.6% MMLU Pro).
What It Means for Developers and Product Teams
Three practical takeaways. First, Apache 2.0 eliminates the “are we allowed” question: embedding, modifying, and redistributing the models requires signing nothing, and what you save is weeks of procurement and legal review, not days. Second, the 26B MoE is the most product-relevant of the bunch: near-31B quality at roughly 4B-active-parameter cost, well suited to high-throughput features and serviceable from a single H100. Third, the edge variants give offline and privacy-sensitive contexts — field devices, no-connectivity environments, data that must not leave the device — a credible option.
The installed base is worth noting too: the Gemma family has passed 400 million cumulative downloads with more than 100,000 community variants, and real deployments range from INSAIT’s Bulgarian-language BgGPT to Yale’s Cell2Sentence-Scale cancer research. Google also states that Gemma 4 went through the same safety protocols as its proprietary models. For teams picking an open-source baseline in 2026, the shortlist just got shorter.
Sources
- Gemma 4 official announcement — Google Blog
- Welcome Gemma 4: Frontier multimodal intelligence on device — Hugging Face
- Google announces Gemma 4, switches to Apache 2.0 license — Ars Technica
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.