Gemma 3 270M

Released August 14, 2025, Gemma 3 270M is Google DeepMind's smallest open model: 270 million parameters combining 170M embedding parameters with 100M transformer blocks, built for task-specific fine-tuning rather than general chat. It is designed for on-device and edge deployment — quantization-aware training lets it run in INT4 on smartphones, Raspberry Pi-class hardware, and browsers — with a 32K context window and a 256K-token vocabulary that handles rare tokens well. Google positions it as a strong base for developers to fine-tune in minutes on narrow, domain-specific tasks.

Limitations: At 270M parameters it has a low quality ceiling — it is a fine-tuning substrate, not a capable generalist, and can produce empty or degenerate outputs under greedy decoding (non-zero temperature recommended). The 32K context is text-only with no multimodal support. Like all tiny models, it is prone to overfitting on small fine-tuning datasets without careful regularization.

Google-reported: efficient fine-tuning; strong for its size class