HN · HACKER_NEWS// 9d agoMODEL RELEASE

Google drops Gemma 4 open models

Google releases its most advanced suite of open-weight models to date, Gemma 4, built on Gemini 3 technology. The release features native multimodal support for audio and vision, a 256K context window, and performance-tier models that rival much larger systems in complex reasoning and agentic workflows.

// ANALYSIS

Gemma 4 is a strategic pivot for Google, shifting to an Apache 2.0 license to challenge the dominance of Chinese open-source leaders like DeepSeek.

–Native audio, vision, and text support across the entire model family enables true multimodal applications on the edge.
–The flagship 31B model's reasoning leap (89.2% on AIME 2026) suggests "intelligence-per-parameter" is the new frontier.
–Agentic focus with native function calling and "thinking modes" for complex planning tasks.
–256K context window for the 31B model significantly reduces RAG complexity for long-document tasks.
–Inclusion of an MoE variant (26B) provides a high-speed, high-intelligence middle ground for production inference.

// TAGS

gemma-4llmopen-weightsagentmultimodalgoogleopen-source

DISCOVERED

9d ago

2026-04-02

PUBLISHED

9d ago

2026-04-02

RELEVANCE

10/ 10

AUTHOR

jeffmcjunkin