gemma-3-1b-it

Gemma 3 1B IT (google/gemma-3-1b-it) is Google DeepMind's smallest instruction-tuned model in the Gemma 3 family, released in March 2025. Built on the same research stack as Gemini, it is a dense decoder-only Transformer with Grouped-Query Attention (GQA), alternating local sliding-window and global self-attention layers, and knowledge-distillation training on roughly 2 trillion tokens. Unlike the 4B, 12B, and 27B variants, the 1B model is text-only (no vision encoder), supports a 32K-token context window, and is optimized for edge, mobile, and resource-constrained deployments.

The instruction-tuned variant targets conversational use, instruction following, structured outputs, and function calling. Official benchmarks show strong instruction adherence (IFEval 80.2) and solid grade-school math (GSM8K 62.8), while more demanding reasoning and knowledge tasks remain limited at this scale (MMLU-Pro 14.7, GPQA-Diamond 19.2, LiveCodeBench 1.9). Quantization-aware training checkpoints (INT4, FP8) further reduce memory footprint for on-device inference.

Gemma 3 1B IT is released under the Gemma Terms of Use with open weights on Hugging Face. It is well suited for lightweight chatbots, on-device assistants, and high-throughput text generation where latency and memory matter more than frontier reasoning depth.