Google Gemma 4 Goes Apache 2.0 — The Open Model War Just Escalated
Google just dropped Gemma 4 and the most important change isn't the model itself — it's the license. Apache 2.0. No asterisks, no fine print, no "Gemma Use Policy" that enterprises need lawyers to parse. After years of restrictive open-weight licensing, Google is saying: take it, ship it, we don't care.
The technical specs are solid but not shocking: MoE architecture with 26B total parameters but only 3.8B active per token, meaning it runs on consumer hardware while punching above its weight. Native function calling, 140+ languages, video and audio inputs. The smaller 2B and 4B variants are designed for edge deployment — phones, laptops, IoT devices. This is Google explicitly targeting the local-first agent stack that PrismML and Ollama have been building toward.
What makes Gemma 4 matter for agents isn't any single benchmark. It's that every agent framework — LangChain, CrewAI, AutoGen — can now embed a commercially unrestricted Google model as their default backbone without worrying about license compliance. That changes the economics of the entire ecosystem. Previously, if you wanted truly open weights for production agents, your options were Llama (Meta's license), Qwen (Chinese model, geopolitical concerns), or Mistral (EU, smaller). Now Google is in the mix with zero strings attached.
1482 points on Hacker News in hours. That number tells you everything about how much pent-up demand existed for a major lab to commit to real open-source AI. The model is available now on Hugging Face, Kaggle, Ollama, Google AI Studio, and AI Edge Gallery.
https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
← Back to all articles
The technical specs are solid but not shocking: MoE architecture with 26B total parameters but only 3.8B active per token, meaning it runs on consumer hardware while punching above its weight. Native function calling, 140+ languages, video and audio inputs. The smaller 2B and 4B variants are designed for edge deployment — phones, laptops, IoT devices. This is Google explicitly targeting the local-first agent stack that PrismML and Ollama have been building toward.
What makes Gemma 4 matter for agents isn't any single benchmark. It's that every agent framework — LangChain, CrewAI, AutoGen — can now embed a commercially unrestricted Google model as their default backbone without worrying about license compliance. That changes the economics of the entire ecosystem. Previously, if you wanted truly open weights for production agents, your options were Llama (Meta's license), Qwen (Chinese model, geopolitical concerns), or Mistral (EU, smaller). Now Google is in the mix with zero strings attached.
1482 points on Hacker News in hours. That number tells you everything about how much pent-up demand existed for a major lab to commit to real open-source AI. The model is available now on Hugging Face, Kaggle, Ollama, Google AI Studio, and AI Edge Gallery.
https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
Comments