MUM explained: Google's multitask unified model and what it means for search

Fibvo editorial · 1 May 2023 · 5 min read

Announced at Google I/O in May 2021, MUM (Multitask Unified Model) represented a further leap beyond BERT, described by Google as roughly a thousand times more powerful and trained across 75 languages simultaneously. MUM's headline capability was understanding information across different formats — text and images together — and across languages, in a single unified model.

What makes MUM different from BERT

BERT was trained to understand text deeply within a single language context. MUM was built to be multimodal and multilingual from the ground up, meaning it can, in principle, understand a question asked partly through an image and partly through text, and draw on information published in a different language than the query itself.

• Multimodal: understanding information across text, images and potentially video together.
• Multilingual: trained across 75 languages simultaneously, allowing it to draw on relevant information regardless of the language it was originally published in.
• Multitask: capable of performing several understanding tasks at once, rather than needing separate models for each.

Google's own example

Google demonstrated MUM with a hypothetical query: someone who has hiked Mount Fuji wanting to know if the same boots would work for Mount Adams, potentially combining a photo of their boots with a text question. Answering that well requires understanding the image, the implicit comparison being made, and detailed information about both mountains — a task well beyond simple keyword matching.

MUM's real significance is not any single feature — it is the shift toward search understanding a question the way a knowledgeable person would, across formats and languages, rather than as isolated text.

Where MUM's influence actually shows up

MUM's capabilities feed into several visible features: related search suggestions that go beyond simple keyword variation, Google Lens' ability to understand more complex visual queries, and the broader language understanding underpinning AI Overviews. It is less a single, isolated ranking factor and more infrastructure powering several search experiences at once.

What this means for content strategy

Content that combines clear text with well-described, relevant images and video, rather than treating them as separate, disconnected elements, is better positioned for a search system built to understand information across formats. Alt text, captions and surrounding context that genuinely describe and relate to accompanying visuals help this multimodal understanding, beyond their accessibility benefits covered in our image SEO guide.

• Write alt text and captions that genuinely describe what an image shows, not just a keyword.
• Ensure video content includes accurate transcripts and descriptions.
• Keep text and visual content on a page thematically coherent, rather than stock imagery unrelated to the surrounding text.

Frequently asked questions

If you want a review of how well your visual and text content work together for both readers and modern search systems, reach out to our team.

---

Written by Fibvo — Bengaluru and Erode, India — working with clients in other time zones. Questions, or think we have it wrong? Tell us.