Google reveals an open code Gemma 3 with 128K context window

Even as large language and models of reasoning remain popular, organizations are increasingly turning to smaller models to perform AI processes with less energy and cost problems.

While some organizations distillated larger models of smaller versions, model providers like Google Continue to launch small language models (SLM) as an alternative to large language models (LLMS), which may cost more to work without sacrificing performance or accuracy.

With this in mind, Google has released the most version of its small model, Gemma, which has extended context windows, larger parameters and more multimodal reasoning options.

Gemma 3 that has the same processing power as greater Models Gemini 2.0It remains best used by smaller devices such as phones and laptops. The new model has four sizes: 1B, 4B, 12B and 27B parameters.

With a larger context window of 128K tokens – as opposed to Gemma 2 If there was a context window of 80K – Gemma 3 could understand more information and complex requests. Google updates GEMMA 3 to work in 140 languages, analyze images, text and short videos and call a support feature to automate tasks and agents.

Gemma gives a strong performance

To further reduce the cost of calculations, Google has introduced quantum versions of GEMMA. Think about quantum as compressed models. This happens through the process of “reducing the accuracy of numerical values ​​in the weight of the model” without sacrificing accuracy.

Google said that Gemma 3 “provides state-of-the-art performance for its size” and superior to the leading LLMs like Llama-405B, Deepseek-V3 and O3-Mini. Gemma 3 27b, in particular, came second in Deepseek-R1 in the ELO chat tests tests. It was covering Deepseeksmaller model, Deepseek V3, OPENAI‘is 3-mini, Meta‘s llama-405b and Mistral Big.

By quantizing GEMMA 3, users can improve performance, launch the model and build applications “that can fit on a host of single GPU and Tensor Processing Unit (TPU).”

Gemma 3 integrates with developer tools such as hugging face transformers, Ollama, Jax, Keras, Pytorch and others. Users can also have access to Gemma 3 via Google Ai Studio, hug or kaggle. Companies and developers can request access to Gemma 3 API via AI Studio.

Shield Gemma for security

Google said it has built safety protocols in Gemma 3, including an image safety check called Shieldgemma 2.

“The development of GEMMA 3 included extensive data management, aligning our safety policies through fine -tuning and stable comparison evaluations,” Google wrote in a blog post. “While the in -depth testing of more capable models often informs our assessment of less capable, the improved effectiveness of the GEMMA 3 STEM caused specific evaluations focused on its potential for abuse in the creation of harmful substances; The results show a low -risk level. “

Shieldgemma 2 is a 4B image safety check, built on the GEMMA 3 Foundation. It finds and does not allow the model to react with images containing sexually explicit content, violence and other dangerous materials. Users can customize Shieldgemma 2 to meet their specific needs.

Small patterns and distillation in the rise

Ever since Google first released Gemma in February 2024Slm have seen Increase in interest ratesS Other small models such as Microsoft’s Phi-4 and Mistral Small 3 Indicate that businesses want to build applications with models as powerful as LLM, but it is not necessary to use all the breadth of what LLM is capable of.

Businesses also began to turn to the fewer LLM versions, which they prefer by distillation. To make it clear, Gemma is not a distillation of Gemini 2.0; He is more trained with the same set of data and architecture. Distilled model learns from a bigger model that Gemma doesn’t do.

Organizations often prefer to meet certain cases of model use. Instead of having LLM like the O3-Mini or Claude 3.7 Sonnet to a simple code editor, a smaller model, whether SLM or distilled version can easily perform these tasks without overcoming a huge model.