Google has released Gemma 4

Google just released Gemma 4, and for once the open-source AI community has something to celebrate that isn't Chinese! The new family consists of four open-weight models spanning from phones to workstations: the edge-oriented Effective 2B and 4B variants, a 26B Mixture-of-Experts, and a 31B Dense model. Built on the same research foundation as Gemini 3, the release is Google's most important move yet in the open-weight space.

At the time of the writing of this Story, the 31B model currently sits at position 3 among all open models on Arena AI's text leaderboard, and the 26B variant secures position 6, while outcompeting models 20 times its size. The 26B MoE achieves 88% on AIME 2026 with only 3.8B active parameters during inference [LushBinary], making it one of the most efficient reasoning models available. On the coding front, the 31B model jumped from a Codeforces ELO of 110 (Gemma 3) to 2150 - a generational leap by any reasonable measure. This kind of improvement isn't usually seen between model generations from the same family.

Furthermore, Gemma 4 supports text generation across more than 140 languages. For institutions and developers outside the English language bubble, it's often a deciding factor. A government or NGO or any other entity can now take the model, fine-tune it on their own data, deploy it on their own infrastructure, and never touch a cloud API. We've stated this before, but that combination of multilingual capability and data sovereignty is something no proprietary model can offer by design.

Google released the model, and NVIDIA optimized it to run efficiently on local hardware. Early testing by Hugging Face showed the models performed so well out of the box that finding meaningful fine-tuning use cases was difficult. From day one, the ecosystem fully supported the release across both local and serving stacks, including llama.cpp, Ollama, MLX, vLLM, Unsloth, and transformers.

While Gemma 4 may not dethrone DeepSeek or Qwen at the top of the OS AI leaderboards, Google’s contribution stands out in its breadth. The model family still offers flexibility across a wide range of devices, from smartphones to workstations, and is well-suited for local fine-tuning and deployment, while supporting a lot of languages, making it an appealing option for institutions outside the English speaking world. In the open-source space, versatility and accessibility are often just as valuable (if not more!) as benchmark performance.