Gemma 4 — Google DeepMind open models family (Apache 2.0)

TL;DR

Google DeepMind announces Gemma 4, a new family of open-weights models released under Apache 2.0 on April 2, 2026. The lineup spans four sizes — E2B and E4B edge variants, a 26B Mixture-of-Experts (with ~3.8B active per token), and a 31B Dense flagship — all derived from the same research stack that produced Gemini 3. The release is pitched as advanced reasoning and agentic workflows that run on user-owned hardware, with multimodal input (text + image, plus native audio on the small edge variants) and 140+ language support.

Key claims

Gemma 4 is released under Apache 2.0, a more permissive license than prior Gemma generations [tweet body].
The family is built from the same research and technology as Gemini 3 [tweet body].
Four sizes are shipped: E2B, E4B, 26B MoE (~3.8B active), and 31B Dense [tweet body].
Models are targeted at advanced reasoning and agentic workflows runnable on user hardware [tweet body].

Method

This is a product-announcement tweet; no methodology is disclosed in the post itself. The detailed model architecture, training recipe, evaluations, and license rationale live in the accompanying Google Developers blog post and the Google Open Source blog post, both also dated April 2, 2026. Per third-party coverage, the edge variants add native audio understanding (a first for the Gemma family), and the 31B Dense and 26B MoE are positioned as workstation / consumer-GPU class respectively. None of this is in the filing artifact — see the linked posts for primary technical detail.

Results

The tweet itself does not include benchmark numbers. Third-party coverage cites a 31B Dense score of 89.2% on AIME 2026 and a #3 ranking on Arena AI among open models at release, but these claims are not in the artifact filed here and are not independently verified by this page.

Why it’s interesting

This is the next-generation open-weights flagship from a frontier lab, and the license shift from Gemma’s prior bespoke terms to Apache 2.0 is the most material change for downstream adoption — Luma’s prior filings on Qwen and Mistral foundation releases all turn on Apache 2.0 being the default open license. Worth tracking against Trinity Large: An Open 400B Sparse MoE Model (open 400B sparse MoE) and Mistral Small 4 119B (instruct + reasoning + Devstral unified MoE) (Mistral Small 4 119B unified MoE) as the open ≥10B-activated-MoE cohort, and against Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters / MiniMax-M2.5 as the small-edge-to-frontier open lineup. The “Gemini 3 derived → open under Apache 2.0” pipeline is the same playbook that Tencent used with HunyuanImage 3.0 (frontier-internal-stack → open release with full report) — see HunyuanImage 3.0 Technical Report.