Gemma 4 was released on April 2, 2026, by Google DeepMind. It is the fourth generation of the Gemma open-source model family, introducing native multimodal understanding (text, image, video, audio), a Mixture of Experts architecture, and 256K context windows.
Below you'll find the complete timeline of the Gemma model family, key milestones, and what changed in each release.
Google's first open-source model family. Released in 2B and 7B sizes. Text-only, supporting English and a limited set of languages. Built on Google's Gemini research but optimized for open distribution.
Incremental update with improved instruction tuning, better safety alignment, and minor performance improvements across benchmarks.
Major upgrade introducing 9B and 27B parameter models. Significant improvements in reasoning, code generation, and multilingual support. Introduced knowledge distillation techniques for smaller models.
Added multimodal capabilities (text + image). Expanded to 4 model sizes (1B, 4B, 12B, 27B). Introduced 128K context window and improved multilingual support for ~30 languages.
The most significant release. Four variants: E2B (2B), E4B (4B), 26B A4B (MoE with 128 experts), and 31B Dense. Native support for text, image, video, and audio. 256K context window. 140+ languages. Built-in function calling and agentic capabilities. Apache 2.0 license.
Gemma 3 supported text and images. Gemma 4 adds native video and audio understanding within the same unified model — no separate encoders needed.
The new 26B A4B variant uses 128 expert networks, activating only 4B parameters per inference. This delivers large-model quality at small-model compute cost.
The 26B and 31B models double the context window from 128K to 256K tokens, enabling processing of entire codebases or book-length documents.
Built-in function calling and structured JSON output enable autonomous tool use, multi-step planning, and integration with external services.
Expanded from ~30 languages in Gemma 3 to 140+ languages, with strong performance verified on the MMMLU multilingual benchmark (85.2%).
Major score jumps: AIME 2026 (89.2%), LiveCodeBench v6 (80%), GPQA Diamond (84.3%). These represent substantial gains over Gemma 3's already competitive results.
Since its April 2, 2026 release, Gemma 4 has been available on all major model platforms:
Gemma 4 was released on April 2, 2026, by Google DeepMind. All four model variants (E2B, E4B, 26B A4B, and 31B) were made available simultaneously.
As of April 2026, yes. Gemma 4 is the latest generation of the Gemma model family. Google has not announced a Gemma 5 release date.
Google has been releasing major Gemma updates roughly every 6-12 months: Gemma 1 (Feb 2024), Gemma 2 (Sep 2024), Gemma 3 (Mar 2025), Gemma 4 (Apr 2026). Minor updates and patches occur more frequently.
Google has not officially announced Gemma 5. Based on the ~6-12 month release cadence, a next-generation model could be expected in late 2026 or 2027, but this is speculation.
Yes. Previous Gemma versions remain available on Hugging Face, Ollama, and other platforms. However, Gemma 4 offers significant improvements across all metrics. For new projects, Gemma 4 is recommended.
Gemma 4 is built on the same research and technology behind Gemini but is released as a fully open-source model under Apache 2.0. Gemini remains Google's proprietary offering available via API.
pages.release-date.releaseDatePage.faq.items.6.a
pages.release-date.releaseDatePage.faq.items.7.a
pages.release-date.releaseDatePage.faq.items.8.a
pages.release-date.releaseDatePage.faq.items.9.a
Gemma 4 is available now. Try it online, deploy it locally, or explore the model variants.