Gemma 3 270M

2025-08-14 · Publisher : Google · Domain : Artificial Intelligence

Google releases Gemma 3 270M, an open model with 270 million parameters designed for fine-tuning on specific tasks, very energy-efficient and able to follow instructions out of the box.

Google releases Gemma 3 270M, an open model with 270 million parameters designed for fine-tuning on specific tasks, very energy-efficient and able to follow instructions out of the box. ### What changes - Compact architecture: 170 million embedding parameters (256,000-token vocabulary) and 100 million for the transformer blocks. - Follows instructions out of the box (IFEval benchmark) and targets classification, extraction and routing after fine-tuning. - Very efficient: the INT4-quantized model uses 0.75% of a Pixel 9 Pro's battery for 25 conversations, according to internal tests. - QAT checkpoints available for INT4 with minimal loss. ### Specifications - Pre-trained and instruction-tuned versions. - Not designed for complex conversations: it is a base to specialize. - Availability: open weights, runnable locally or on device. ### Pricing Open model, with no usage pricing of its own. ### Benchmarks IFEval: Google reports a level never seen before at this size, shown as a chart (figure not included in the text). ### Key takeaways Useful for high-volume, well-defined tasks, where cost, latency and privacy matter more than versatility.

Tags : Open Source, LLM, Google, Gemma, Embarqué