FLAN-UL2 Models by Google DeepMind

Google DeepMindApache 2.0Open source

2 models2022Up to 2k ctxFrom $0.185/1M input

Last refreshed 2026-05-22. Next refresh: weekly.

Details

ResearcherGoogle DeepMind

LicenseApache 2.0OSI-approved

Commercial useCommercial use: permitted

Models2

Released2022

Max context2k

Links

Website HuggingFace

About

The FLAN-UL2 family of large language models is an advancement of the original UL2 model, leveraging the T5 architecture. Its most notable innovation is the significant expansion of the receptive field from 512 to 2048, which greatly enhances its effectiveness for few-shot in-context learning. Unlike the UL2 model, FLAN-UL2 simplifies operations by removing the need for mode switch tokens during inference and fine-tuning. The model is fine-tuned using the "Flan" prompt tuning method and a specially curated dataset, boosting its few-shot learning abilities. Available in multiple sizes, FLAN-UL2 models can be accessed through their GitHub repository for further exploration and utilization 138.

Decision facts

Best fit: General model comparison
Capability starting point: Flan-UL2 on IBM Watsonx with 2k context
Lowest tracked input: Flan-UL2 on IBM Watsonx · $0.185/1M · IBM watsonx
Closest related family: T5Gemma

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

2 in view

Flan-UL2Current

Use when the workload needs 2k context and 20B parameters.

2022-102k context20B parameters

Flan-UL2 on IBM WatsonxCurrent

Use when the workload needs 2k context and 20B parameters.

2022-102k context20B parameters

Current FLAN-UL2 variants with use-when guidance and lifecycle status
Model	Use when	Released	Signals	Status
Flan-UL2	Use when the workload needs 2k context and 20B parameters.	2022-10	2k context20B parameters	Current
Flan-UL2 on IBM Watsonx	Use when the workload needs 2k context and 20B parameters.	2022-10	2k context20B parameters	Current