Summary: Google DeepMind shipped Gemma 4 on April 2, 2026 under Apache 2.0, replacing the previous custom Gemma terms. The family includes E2B and E4B edge models, a 26B MoE with about 3.8B active parameters, and a 31B dense model, covering more than 140 languages and up to 256K context. Hugging Face’s day-one evaluation puts the 31B at 85.2% on MMLU Pro.
Google DeepMind shipped Gemma 4 on April 2, 2026, a full year after the last major overhaul of its open model family. The headline change has nothing to do with capability and everything to do with licensing: every model in the lineup now arrives under Apache 2.0, and the custom Gemma terms that covered the first two generations are gone. For the first time, a Google open model sits inside a standard open-source license — no extra usage restrictions, no vendor terms that can shift unilaterally, and nothing unfamiliar for a corporate legal team to review.
A single release spans the whole deployment spectrum: the E2B and E4B edge models, a 26B mixture-of-experts variant that keeps only about 3.8B parameters active at inference, and a 31B dense model — all grown from the research and technology behind Gemini 3, natively covering more than 140 languages and reaching as much as 256K of context. Google’s pitch runs one line: its most capable open models to date, built for advanced reasoning and agentic workflows.
The License Move: Custom Terms Out, Apache 2.0 In
Commercial use was already free across the previous two generations, but it came under Google’s own license: users accepted attached terms, and naming plus usage stayed restricted. Individual developers hardly felt the friction; enterprise platform teams could not say the same, since a bespoke license brings legal review, vendor risk assessment, and the standing chance that the terms change under you. In Ars Technica’s framing, the switch to Apache 2.0 was the core news of this release, not a garnish on the marketing.
Timing carries weight here too. Our opening-of-year outlook from January was still tracking reports that Meta might take its Avocado frontier model closed-source. Three months on, Google has pushed the other way with a plain Apache 2.0 grant. The open-weights camp is splitting apart, and Gemma 4 plants Google firmly on the fully open side.
Four Sizes, from a Phone to One H100
The design idea is a single research foundation serving several deployment targets:
- E2B (about 2B effective parameters) and E4B (about 4.5B effective): built for phones, the Raspberry Pi, and NVIDIA Jetson Orin Nano, running offline with latency near zero, and tuned jointly with the Google Pixel team, Qualcomm, and MediaTek
- 26B MoE: built around latency, carrying 26B total parameters while keeping roughly 3.8B active at inference — a balance struck between quality and cost
- 31B dense: the quality pick, which Google positions as a strong base for fine-tuning
On hardware, the two larger models fit on a single 80GB H100 even at bf16 without quantization, while quantized builds run on consumer GPUs. Context runs 128K on the edge models and as much as 256K on the larger pair.
Scores and Leaderboard Standing
Hugging Face’s day-one evaluation of the instruction-tuned models reads as follows: the 31B puts up 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 comes back with 82.6% MMLU Pro, 88.3% AIME 2026, 82.3% GPQA Diamond, and 77.1% LiveCodeBench. Set against the previous Gemma 3 27B, most of these numbers sit clearly higher.
The leaderboard makes the picture even more tangible: as of April 1, the 31B ranked third among open models on the Arena text leaderboard and the 26B sixth, with the MoE variant earning an estimated LMArena score in the region of 1441 on roughly 4B active parameters. The claim that it “outcompetes models 20x its size” is marketing talk — but an open model reaching the open-source top tier on edge-class active parameters is economics that genuinely holds up.
Multimodal Input and Agent-Ready Behavior
Capability-wise, every size takes image and video natively at variable resolutions, with OCR and chart understanding built in; E2B and E4B add native audio input on top. For agentic work, the models cover GUI element detection, object detection returned as JSON bounding boxes, multimodal function calling, and a thinking mode — the perception needed to “read the screen, click the button, call the tool” is effectively part of the weights.
The ecosystem was in place on day one: weights are downloadable from Hugging Face, Kaggle, and Ollama, and transformers, llama.cpp, MLX, Mistral.rs, TRL, Unsloth, and Vertex AI all supported them from release day. On Android, Gemma 4 drives Agent Mode in Android Studio and stays forward-compatible with a future Gemini Nano 4 through AICore. An experimental DiffusionGemma (26B base, 3.8B active) generates text by discrete diffusion, passing 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).
The Practical Read for Developers and Product Teams
Three points worth taking away. Apache 2.0, first, retires the “are we allowed” question: embedding, modifying, and redistributing the models now requires no signatures, and the saving is counted in weeks of procurement and legal review rather than days. Second, the 26B MoE is the strongest product play of the four: near-31B quality for roughly 4B-active-parameter cost, comfortable on a single H100 and a natural fit for high-throughput features. Third, the edge variants hand offline and privacy-sensitive contexts — field devices, no-connectivity environments, data that must not leave the device — a credible option.
The footprint matters as well: the Gemma family has crossed 400 million cumulative downloads and spawned more than 100,000 community variants, with real deployments stretching from INSAIT’s Bulgarian-language BgGPT to Yale’s Cell2Sentence-Scale cancer research. Google also notes that Gemma 4 went through the same safety protocols as its proprietary models. For teams choosing an open-source baseline in 2026, the list of first picks is shorter now.