Google releases Gemma 4, a family of open models built off of Gemini 3
agents gemini gemma google openai open-source
| Source: Engadget | Original article
Google has unveiled Gemma 4, a quartet of open‑weight language models that inherit the architecture and training advances of its flagship Gemini 3 system. The models, released under an Apache 2.0 licence, span from a 2 billion‑parameter variant designed for smartphones to a 13 billion‑parameter version aimed at workstation‑class workloads. All four support multimodal inputs and are marketed as “responsible AI” tools that can be fine‑tuned for commercial or research use without the licensing constraints of Google’s proprietary offerings.
The launch matters because it translates the most sophisticated research from Gemini 3—Google’s latest proprietary LLM—into a publicly available stack. By delivering higher “intelligence‑per‑parameter” than the preceding Gemma 3, the new family narrows the performance gap between closed‑source giants and community‑driven models such as LLaMA 3 or Mistral 7B. For Nordic startups and research labs, the Apache licence removes a major barrier to experimentation, enabling local fine‑tuning on sensitive data and tighter integration with on‑prem hardware. The move also signals Google’s intent to shape the open‑model ecosystem, potentially steering standards around safety mitigations and evaluation metrics.
As we reported on 9 April, Gemma 4’s cost‑efficiency promises to make advanced AI more accessible; the current rollout adds the missing open‑source counterpart to that narrative. The next weeks will reveal how quickly the community adopts the models, whether benchmark results confirm the claimed gains, and how Google positions Gemma 4 alongside its cloud AI services. Watch for announcements of tooling support, such as integration with LM Studio or Google’s Vertex AI, and for any policy updates that could affect commercial deployment in Europe’s regulated markets.
Sources
Back to AIPULSEN