Google DeepMind released Gemma 4 on April 2 under the Apache 2.0 license, the first Gemma release under an OSI-approved open-source license and the most significant move in Google's open-weights strategy to date. Previous Gemma versions shipped under the "Gemma license," which imposed commercial-use restrictions that kept the models out of many production pipelines. Apache 2.0 removes those constraints. The family spans four sizes: E2B and E4B "effective" edge variants, a 26B Mixture-of-Experts, and a 31B dense model that sits at #3 on the open-model Arena leaderboard. All models are natively multimodal (video, image, OCR, chart understanding with variable-resolution input) and agentic (function-calling, structured JSON output, system-instruction support). The edge E2B and E4B variants additionally accept native audio input for speech recognition and understanding. Context windows are 128K on edge and up to 256K on the larger models, with native training across 140+ languages.
The licensing change matters as much as the technical specs. Teams that wanted to build on Gemma 3 and discovered the commercial clauses in the old license, often late in the project, had to choose between rewriting against Llama, Mistral, or Qwen weights, or accepting the restrictions and explaining them to customers. Apache 2.0 is commercially permissive, patent-friendly, and compatible with the vast majority of corporate legal stances. It is what enterprise buyers actually want when they say they want open weights. On capabilities, the 31B dense model's #3 Arena open-model ranking is real, and the 26B MoE at #6 is strong on cost-per-inference. The E2B and E4B edge models are the more novel story. Edge-friendly multimodal with native audio input in a 2-to-4B-effective footprint is the first genuine on-device alternative to the proprietary edge models from Apple and Qualcomm, and context windows of 128K are large enough for real document-processing workloads rather than toy demos. Agentic-native design choices, with function calling and structured JSON as first-class outputs and system instructions supported at the protocol level, also reduce the custom scaffolding that teams have been writing for two years.
The open-weights landscape for April 2026 now looks coherent rather than fragmented. Llama's commercial-use caveats still exist, Mistral's licensing terms vary by model, Qwen is Apache 2.0 but carries perception-of-origin risks for some buyers, DeepSeek is capable but has similar geopolitical considerations. Gemma 4 under Apache 2.0 from a Google-scale lab changes the procurement conversation for enterprises that want open weights without either regulatory exposure or commercial restrictions. The impact on the managed-API business is more interesting than most of the coverage has acknowledged. If you can run a 31B model at home-grade inference cost and get #3-on-Arena quality with native function calling, the economic case for always calling Anthropic's or OpenAI's API weakens for task classes that do not specifically need frontier reasoning. That does not threaten the top labs immediately, because model quality at the frontier is still the differentiator for complex work. It does compress the mid-tier API business, which is where most volume actually lives.
Three concrete moves for builders. First, evaluate Gemma 4 against whatever you are currently using for the mid-tier of your model routing; the Apache 2.0 license removes the old "we cannot ship this in production" blocker, and the capabilities may close the quality gap for bulk tasks. Second, the E2B and E4B edge variants are worth prototyping against for any workflow where on-device inference would change the product shape, specifically privacy-sensitive data, offline operation, and low-latency interaction. The audio-input support is specifically interesting for voice-first products. Third, the agentic-native design (function calling as first-class, structured JSON, system instructions) means less custom scaffolding for agent deployments. Teams that built their own tool-calling shims for Gemma 3 can delete code. The license change, not the benchmarks, is the detail that changes roadmap conversations with legal and procurement. If you previously argued for a proprietary API because open weights were "not commercially clean," that argument just got weaker.