Flan-T5 XL

Flan-T5 XL

Quick answer: Flan-T5 XL is Google's 3-billion-parameter instruction-tuned encoder-decoder model from the "Scaling Instruction-Finetuned Language Models" paper (Chung et al., 2022). It is open-weight (Apache 2.0) and is the largest Flan-T5 backbone commonly fine-tuned in agent and grounding research, including as the best-performing MindAct backbone in the Mind2Web benchmark.

At a Glance

Historical significance

• Largest widely-used Flan-T5 checkpoint, requiring multi-GPU fine-tuning
• Best-performing fine-tuned backbone in the original Mind2Web (MindAct) results, ahead of Flan-T5 Base/Large

Modern context

• Long surpassed in raw capability by modern decoder-only LLMs
• Still cited as a fully open, fine-tunable reference point in academic agent benchmarks