MUM (Multitask Unified Model)

MUM, the Multitask Unified Model, is an AI model Google announced at I/O in May 2021. Google described it as built on the T5 text-to-text framework, trained across 75 languages and many tasks at once, multimodal (able to understand text and images), and able to generate language as well as understand it. It sits in the line of Google language models between BERT and today’s Gemini models.

How it works

Like BERT, MUM is a Transformer-based model. The differences Google emphasised were scale (it called MUM 1,000 times more powerful than BERT), multitask training, multilingual transfer (learning from sources in one language and applying that knowledge in another) and multimodality. The idea was to help with complex needs that used to take several searches, such as comparing two hikes or preparing for a trip, by drawing together information across languages and formats.

Google’s guide to its ranking systems states that MUM is not currently used for general ranking in Search. It is used for specific applications, such as improving searches for COVID-19 vaccine information and improving featured snippet callouts.

How it’s used in SEO and marketing

• Context: MLforSEO lessons use MUM, with BERT, as an example of how Google combines multiple machine learning models within its wider search system, and of Google’s ambition to rely more on machine understanding of content.
• Journey thinking: MUM was presented around mapping complex tasks, which links to planning content across the whole user journey with the knowledge graph and the EAV model.
• Multimodal content: MUM signalled the direction later realised in Google Lens features and multimodal search, so images and video deserve the same care as text.

Example

Google’s launch example was a hiker asking how to prepare for Mount Fuji after already hiking Mount Adams. A traditional search needs many separate queries about elevation, weather, gear and trails. MUM was designed to understand the comparison, draw on Japanese-language sources, and point to useful information across those sub-topics, an early version of the multi-step behaviour now seen in AI search.

Limitations and things to watch

• There is no way to “optimise for MUM” specifically, and Google says it is not a general ranking system.
• Much of what MUM promised has since been delivered by newer Gemini models, AI Overviews and AI Mode, so treat older MUM commentary as historical context.
• Public detail on MUM is limited to Google’s announcements; avoid claims that go beyond them.

Learn it on MLforSEO

• Image SEO in 2026: The Complete Guide to Image Optimisation, Advanced Tactics and Multimodal Search
• What Is Implicit User Feedback in SEO? How Behavioural Signals Shape Search Rankings
• What Is the EAV Model in SEO? A Practical Framework for Semantic Keyword Research
• Introduction to Machine Learning for SEO (course)
• Semantic ML-enabled Keyword Research (course)