Google Drops Gemma 4: Four Open-Weight Models With 256K Context, Multimodal Input, and Top-3 Arena Ranking
Google released Gemma 4 on April 2, 2026 — four open-weight models (E2B, E4B, 26B MoE, 31B Dense) built from the same research as Gemini 3. The 31B ranks #3 among all open models on the Arena AI leaderboard. Every size supports image, video, and audio input out of the box.
Original sourceGoogle DeepMind released Gemma 4 on April 2, 2026, describing it as "byte for byte, the most capable open models" to date. The family spans four model sizes — Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture-of-Experts (MoE), and a 31B Dense — all released under Apache 2.0 and available on Hugging Face and Google AI Studio.
Every Gemma 4 model supports multimodal input out of the box: image understanding at variable resolution, video comprehension up to 60 seconds, and audio input for speech and translation. All models include a 256K token context window, native function calling, a configurable reasoning/thinking mode, and structured output support — features previously confined to closed commercial models.
On benchmarks, the 31B model currently ranks #3 among all open-weight models on the industry-standard Arena AI text leaderboard. On AIME 2026 (math competition benchmarks), the 31B scores 89.2% compared to Gemma 3 27B's 20.8% — a 4x improvement in one generation. The 26B MoE achieves 88.3% with only 3.8B active parameters, making it extremely efficient for deployment.
The release cements Google's position in the open-weight race alongside Meta's Llama 4 and Alibaba's Qwen family, and the Apache 2.0 license ensures commercial usability without restrictions.
Panel Takes
“256K context and native function calling at 26B parameters changes what's possible with local deployments. The MoE efficiency especially — 3.8B active params for 88% AIME accuracy is wild.”
“Google has a pattern of releasing impressive open models and then under-investing in the ecosystem. Gemma 3 had similar benchmarks and still got eclipsed in real-world adoption. Benchmarks aren't deployment.”
“When open models built from the same research as Gemini 3 are available under Apache 2.0, the divide between frontier and accessible AI collapses. This is a landmark release for the field.”