---
title: "The Firefox AI Platform"
license: CC-BY-2.0
license_url: http://creativecommons.org/licenses/by/2.0/be/deed.en
license_note: "The FOSDEM video directory https://video.fosdem.org/2025/ub5230/ states: This work is licensed under the Creative Commons Attribution 2.0 Belgium Licence, with the deed at http://creativecommons.org/licenses/by/2.0/be/deed.en. Confirmed 2026-10-18. The VTT file is the publisher caption of that same recording. This note is an original summary, not a transcript."
source: https://archive.fosdem.org/2025/schedule/event/fosdem-2025-4236-the-firefox-ai-platform/
source_license: "CC BY 2.0 Belgium"
attribution: "Summary by iinventedlight (AI-assisted). Source: Tarek Ziadé, The Firefox AI Platform, FOSDEM 2025 Mozilla devroom (CC BY 2.0 Belgium)."
date: 2026-10-18
status: published
category: talks
tags: assistants, local inference, personal data, browsers
checked_on: 2026-10-18
sources:
  - id: S1
    title: "The Firefox AI Platform"
    authors: "Tarek Ziadé"
    url: https://archive.fosdem.org/2025/schedule/event/fosdem-2025-4236-the-firefox-ai-platform/
    accessed: 2026-10-18
    license: CC BY 2.0
    attribution: "Tarek Ziadé, The Firefox AI Platform, FOSDEM 2025 Mozilla devroom (CC BY 2.0 Belgium)."
  - id: S2
    title: "Publisher captions (VTT) for the same recording"
    authors: "FOSDEM"
    url: https://video.fosdem.org/2025/ub5230/fosdem-2025-4236-the-firefox-ai-platform.vtt
    accessed: 2026-10-18
    license: CC BY 2.0
    attribution: "FOSDEM, captions for Tarek Ziadé, The Firefox AI Platform (CC BY 2.0 Belgium). License stated on https://video.fosdem.org/2025/ub5230/."
---
# The Firefox AI Platform

License: [CC BY 2.0 Belgium](http://creativecommons.org/licenses/by/2.0/be/deed.en). Source: https://archive.fosdem.org/2025/schedule/event/fosdem-2025-4236-the-firefox-ai-platform/

## Summary
Tarek Ziadé (Mozilla Firefox AI/ML), speaking at FOSDEM on 2 February 2025, explains a browser inference path that can run without sending the page to a remote model. Firefox translations, started around 2019 with the Bergamot project and Marian machine translation, already did this with small models, WebAssembly, and native backends. Newer ideas on the list include image-to-text, named-entity recognition, tab grouping, semantic search, and speech. The runtime they shipped is Transformers.js, the Hugging Face JavaScript port, on ONNX Runtime Web, so the same code can use WebAssembly or WebGPU and a large catalog of ONNX models. He says it landed in Firefox 133. Inference sits in its own process, away from the page, and downloaded models go in IndexedDB so origins can share them. The first built-in use is alt text in PDF.js: an image encoder plus a small decoder, on the order of a couple hundred megabytes. A web-extension trial API, behind a preference he places in Firefox 134, lets an extension create an engine and ask for a summary. He is clear that this permission is Mozilla's trial surface, not a W3C standard, and he thinks shipping a web-facing model API more widely is early. This note follows the captions. The talk mentions docs and a chat channel without URLs that the captions state cleanly, so those are left out.

## Why it is here
Assistant notes talk about models as if they always live on someone else's server. This talk is a concrete description of local inference in a browser, which is one way to keep a document on the machine while still using a model.