Flan-T5

An instruction-tuned evolution of Google's T5 model that excels at zero-shot tasks through massive-scale multitask fine-tuning.

Flan-T5 is a high-performance encoder-decoder transformer trained on the FLAN collection (1,800+ tasks). By fine-tuning the original T5 architecture on a diverse mixture of instructions, Google researchers improved performance across benchmarks like MMLU and BBH by double-digit margins. It scales from Small (80M parameters) to XXL (11B parameters), offering a commercially viable alternative to closed-source models for summarization, translation, and reasoning. Its primary strength lies in its ability to follow complex prompts without requiring specific task-based training data.

https://huggingface.co/google/flan-t5-xxl