Gemma 3 270M

google/gemma-3-270m is a text-generation model from Google DeepMind, the smallest variant in the Gemma 3 family of lightweight open models. It is trained on 6 trillion tokens (knowledge cutoff August 2024) and supports a 32K-token context window.

Despite its compact size, the 270M model delivers competitive performance on a range of reasoning and language understanding benchmarks. Key results for the pre-trained (PT) variant include HellaSwag (10-shot) 40.9%, BoolQ (0-shot) 61.4%, PIQA (0-shot) 67.7%, and ARC-e (0-shot) 57.7%. The instruction-tuned (IT) variant achieves IF Eval (0-shot) 51.2%, PIQA (0-shot) 66.2%, and WinoGrande (0-shot) 52.3%.

Architecturally, Gemma 3 employs an increased ratio of local to global attention layers with short local spans to reduce KV-cache memory usage, enabling efficient handling of long sequences. The model was trained using JAX and ML Pathways on Google TPU hardware.

As a hosted API on gigarouter, this model provides a low-latency, cost-efficient text generation endpoint suitable for systems where response speed and resource efficiency are priorities, while retaining the reasoning and instruction-following capabilities demonstrated by its benchmark scores.