On April 2, 2026, Google DeepMind released Gemma 4, a family of multimodal open-weight models that range from phone-sized "effective 2B" models to a 31B dense model, all under the Apache 2.0 license. The license is the headline: earlier Gemma generations shipped under Google's custom Gemma Terms of Use, while Gemma 4 uses a standard open-source license with no extra usage policy to review. Combined with strong benchmark results for its size, Gemma 4 became one of the most practical choices for developers who need capable models on laptops, edge devices or a single GPU.

Key Facts#

  • Release: April 2, 2026, with weights on Hugging Face, availability on Google Cloud, and support in local runtimes such as Ollama and llama.cpp.
  • Launch sizes: E2B (2.3B effective parameters, 5.1B including embeddings), E4B (4.5B effective, 8B including embeddings), a 26B mixture-of-experts model with about 4B active parameters, and a 31B dense model.
  • Context windows: 128K tokens for E2B and E4B, and 256K tokens for the 26B and 31B models.
  • Modalities: text, image and video input for all models, plus audio input for E2B and E4B. Output is text.
  • License: Apache 2.0. Google's Gemma Terms of Use page now states that it covers earlier models and points Gemma 4 users to a separate Apache 2.0 license.
  • Languages: support for more than 140 languages, according to Google.
  • Benchmarks: Hugging Face reports estimated LMArena text scores of 1452 for the 31B model and 1441 for the 26B MoE model. Google's published results list 89.2 and 88.3 percent on AIME 2026 without tools, compared with 20.8 percent for Gemma 3 27B.

What Happened#

Google described Gemma 4 as "byte for byte, the most capable open models" and said it was built from the same research as Gemini 3. The family keeps the Gemma 3 pattern of base and instruction-tuned checkpoints for each size, but adds features aimed at agents and on-device use: a thinking mode, function calling that also works with image input, and more flexible vision processing.

Hugging Face's technical breakdown describes an architecture tuned for efficiency. Layers alternate between local sliding-window attention (512 tokens in smaller models, 1,024 in larger ones) and global full-context attention. The small models use Per-Layer Embeddings, first seen in Gemma 3n, which give each decoder layer a small token-specific signal and let a 5B-parameter model run with a memory footprint closer to 2B. A shared KV cache lets the last layers reuse key and value tensors from earlier layers, which saves memory in long contexts. The vision encoder now preserves aspect ratios and lets you choose an image token budget from 70 to 1,120 tokens, so you can trade detail for speed.

The launch had broad day-one support. Hugging Face listed Transformers, llama.cpp with image input, MLX for Apple silicon, Transformers.js with WebGPU and the Rust engine mistral.rs, and ONNX checkpoints were released for edge and browser use. Ollama's own README now uses ollama run gemma4 as its getting-started example.

Background#

Google launched Gemma in 2024 as the open counterpart to Gemini. Gemma 3, released on March 12, 2025, added image input, 128K context and support for more than 140 languages across 1B to 27B sizes. Gemma 3n followed in June 2025 with the "effective parameter" design for phones and laptops. Throughout, the models used Google's own Gemma Terms of Use, which incorporate a Prohibited Use Policy whose restrictions must also be passed on with derivative models. Legal teams had to review those terms before shipping.

By 2026, the competitive landscape had changed. Alibaba's Qwen models, released under Apache 2.0, had become the most common base for community fine-tunes, and Chinese labs were shipping large models under Apache 2.0 or MIT. Moving Gemma 4 to Apache 2.0 removed a key adoption barrier and put Google's small models on the same legal footing as their main competitors.

Why It Matters for Developers#

The license change simplifies shipping. Apache 2.0 is already on most companies' approved-license lists, so a .NET desktop, mobile or server application that embeds Gemma 4 no longer needs a separate review of model-specific terms. That matters most for on-device scenarios, where you distribute the weights to customers instead of calling an API.

The size range fits real hardware. E2B and E4B target phones, laptops and small edge devices, and they accept audio as well as images. That opens local speech and document scenarios, for example in a .NET MAUI app, without sending data to the cloud. The 26B MoE model activates about 4B parameters per token, so it runs fast on a single GPU while scoring close to the 31B dense model. For most internal assistants, it is the size to benchmark first.

Integration options for .NET are good. During development, Ollama plus OllamaSharp gives you an IChatClient from Microsoft.Extensions.AI. For packaged applications, the ONNX checkpoints open a path to ONNX Runtime, which our guide to local AI with ONNX Runtime, Ollama and Foundry Local covers. The adjustable image token budget is worth exposing in configuration: low budgets are fine for classification, and high budgets help with OCR and dense documents. Our multimodal AI guide shows how to pass images and audio through the same abstraction.

Function calling with image input is also practical for agents. A model can read a screenshot or photo, decide which tool to call and pass structured arguments, all locally. As always, validate tool arguments on the server side and test tool-call accuracy with the quantization you actually deploy, as described in our function-calling guide.

What's Next#

The family kept growing after launch. Hugging Face's updated post documents a Gemma 4 12B "Unified" model that drops separate vision and audio encoders, multi-token prediction drafters for speculative decoding with reported speedups of up to about three times, and DiffusionGemma, an experimental model that generates text by discrete diffusion. Hugging Face reports that DiffusionGemma exceeds 1,100 tokens per second per user at low batch sizes, while trailing the standard 26B model on most accuracy benchmarks.

Adoption data suggests the strategy is working. Hugging Face's summer 2026 report counted Google as the second-largest source of derivative models on the Hub, and Gemma GGUF builds were downloaded about 20.8 million times a month, second only to Qwen. The open question is whether Google will keep future Gemma generations under Apache 2.0. For now, developers can treat Gemma 4 as a low-risk default for local and edge AI.

Sources#