Google DeepMind today releases Gemma 3 270M, a 270-million-parameter open-weights model that the company says is built for task-specific fine-tuning and extreme energy efficiency on edge devices.

The model combines 170 million embedding parameters with 100 million transformer-block parameters, using a 256k-token vocabulary. Google claims the instruction-tuned version scores 51.2% on the IFEval benchmark, outperforming similarly sized models like Qwen 2.5 0.5B Instruct and SmolLM2 135M Instruct. Internal tests on a Pixel 9 Pro show an INT4-quantized build using roughly 0.75% battery over 25 conversations. The model is available on Hugging Face, Ollama, Kaggle, LM Studio and Docker, with quantization-aware-trained INT4 checkpoints.

Google positions Gemma 3 270M as a ‘right tool for the job’ alternative to billion-parameter models, citing Adaptive ML’s work with SK Telecom, which fine-tuned a Gemma 3 4B model for multilingual content moderation and exceeded the performance of larger proprietary models. The company says the 270M variant can be fine-tuned in hours for tasks like sentiment analysis, entity extraction, query routing, unstructured-to-structured text processing, creative writing, and compliance checks, and can run entirely on-device for privacy-sensitive applications.

Reactions online include praise and criticism, with one Google developer advocate joking the model could run in a toaster. But Liquid AI’s Ramin Hasani notes that Google’s benchmark chart leaves out his company’s LFM2-350M, which he says scores 65.12% on IFEval with only a few more parameters. Google releases both pretrained and instruction-tuned checkpoints.