W3C

WebML CG F2F TPAC 2026 – 30 October 2026

30 October 2026

Attendees

Present
Ali_Spivak, Andrew_Wafaa, Anssi_Kostiainen, Christian_Liebel, Ehsan_Toreini, Gaurang_Tandon, Isaac_Ahouma, Marijn_Kruisselbrink, Markus_Tavenrath, Mathieu_Giroux-Huppe, Mike_Wasserman, Mikhail_Klimenko, Reilly_Grant, Ruoxi_Ran, Samuel_Dallstream, Sohum_Chatterjee, Sushanth_Rajasankar, Yunhao_Luo
Regrets
-
Chair
Anssi
Scribe
Anssi, anssik

Meeting minutes

Built-in AI APIs overview

Anssi: any implementation status updates from Chrome, Edge or others?
… Kenji has kept Awesome Built-in AI APIs repo updated for applications & demos that make use of these APIs:

webmachinelearning/awesome-built-in-ai

Prompt API

Repository: webmachinelearning/prompt-api

Model selection and availability

Anssi: issue #169

<gb> Issue 169 [Tag Review] - Model selection and availability (by etiennenoel) [tag-tracker]

Anssi: this issue discussed Mozilla's and in part also WebKit's and TAG's concerns
… we made a resolution that address these concerns: "Prompt API specification to require conformant implementations to use a permissively licensed model provided either by the implementation or the underlying platform. As further work, determine what criteria makes sense exactly for such models. (issue #169)"

11 June minutes

Anssi: any questions on how to interpret or implement this resolution?

Markus: implementations will always use different models, just as coding agents give different results
… in the future, allowing custom models through the same APIs could give sites consistency and take advantage of specialized hardware
… for now, the value is in the task APIs, accepting that the models behind them differ
… this ecosystem is not fully deterministic, so the focus should be on the APIs and their behavior

Anssi: in response to MarkusT's comment, there's an issue for letting the user choose a model, see issue #8

<gb> Issue 8 Choose model (by niutech)

Reilly: the ecosystem of models is evolving; the cross-origin storage API takes an approach opposite to built-in AI APIs
… middleground for task-focused APIs, users can pick a model
… interop one is important, benefits to browser vendors to optimize a model to run on the device
… developer choice means they lose control over model performance
… a tradeoff to take

Markus: device classes matter; mobile devices have specialized chips and highly optimized local models, while desktops have more varied hardware

Reilly: on capable systems, browsers might let users opt into a larger, more resource-intensive model for better quality
… sites should not be able to force users to download such a large model

Sushanth: the interop dataset sits beside the specification, so developers can still use the Prompt API for uncaptured use cases
… canonical prompts for common scenarios such as classification or rewriting might tie interoperability more closely to the API
… developers should be encouraged to contribute to the interop dataset

Anssi: the existing resolution still holds, with no further resolution needed

New features under development

Reilly: sampling modes let developers select points on a spectrum from creative to predictable output
… raw temperature and top-k parameters pose interoperability problems because models respond differently
… fixed modes can preserve predictable behavior while their model-specific parameters evolve
… thinking effort similarly exposes fixed levels rather than raw model parameters
… a fastest mode lets developers prefer decoding speed without prescribing a particular implementation strategy
… proposals and implementations of these features will be available for deeper review

Anssi: a number of features under development, message format for tool calling, thinking, and multimodality #209

sampling modes #203, thinking effort #222, and fastest mode #219

<gb> Issue 203 Proposal: Introduce Categorical Sampling Modes for the Prompt API (by isaacahouma) [enhancement] [interop]

<gb> Issue 209 Extend the Message format to support tool calling, extended thinking, and multimodality (by nico-martin)

<gb> CLOSED Issue 222 Thinking effort API - includeThoughts (by sushraja-msft)

<gb> Issue 219 Consider adding a "fastest" LanguageModelSamplingMode value (by isaacahouma)

Sushanth: multimodality has interoperability risks, especially for audio when current small models offer little beyond transcription
… image captioning may be better served by a task API; text issues should be addressed before expanding Prompt API multimodality
… audio transcription may belong in the Web Speech API rather than the Prompt API

Mikhail: should multi-token prediction parameters be included as the API exposes more low-level controls?

Reilly: multi-token prediction motivates the fastest mode because it is currently available only with particular sampling modes
… the generic mode avoids baking a particular decoding strategy into the API
… Web Speech is appropriate for streaming transcription; other audio and image uses need more developer experience and evaluation examples
… multimodality extensions are forward-looking and should use feature detection as model capabilities improve

Isaac: we want community feedback, especially from Microsoft, on the proposal's seven sampling levels before shipping

Sushanth: extreme sampling levels can make small language models loop, and it is unclear why developers need seven interoperable levels
… seven levels can work if implementations map them to the nearest supported level and interop tests set expectations

Agenda bashing

Reilly: writing assistance APIs may use smaller expert models, so developers could choose a faster option with fewer presets or languages
… privacy and model-location choices were discussed in an earlier breakout
… the breakout also covered Translator and Language Detector APIs; follow-up questions are welcome

Repository: webmachinelearning/writing-assistance-apis

Sushanth: Translator and Language Detector APIs appear mature and interoperable, though quality evaluation data is still needed
… for writing assistance, adapters can specialize small models without requiring a heavyweight new task API
… a lightweight way to specify a task on the Prompt API could be valuable

Reilly: some small expert models may not understand general prompts and could mistakenly summarize the instructions themselves
… dedicated APIs ensure requests are framed correctly, though task specification deserves further exploration

Translator and Language Detector APIs

Sushanth: dynamic batching now provides the desired parallelism internally, so an API change may not be necessary
… the original goal was to let developers trade throughput for latency

Batched translation

Anssi: issue #85

<gb> Issue 85 not found

Repository: webmachinelearning/translation-api

Sushanth: batched translation #85 is on pause until we learn more

Reilly: batching adds API and semantic complexity, and Firefox's model capabilities should also be considered
… an HTML translation API might inspect a large snippet and batch suitable parts internally
… that could reduce the need for a separate explicit batching API

Ehsan: splitting a page into sections can lose context and produce poor translations

Reilly: the browser can decide how much context a model needs and use HTML semantics to identify parallelizable sections

Mikhail: batching is also useful for independent short items such as user reviews

Reilly: implementations using traditional machine translation may need little context, so browsers should control the optimization

Ehsan: +1 to Sushanth on the status of the Translation API and its potential to move to the WG

Markus: batch translation should avoid multiplying memory use for large inputs and their key-value caches

Reilly: a batch is only a hint that parallelism is available; the browser chooses an amount appropriate for system resources

Proposals for new incubations

Repository: webmachinelearning/proposals

Decisions API

Anssi: Decisions API #20

<gb> Issue 20 Decision Models (by AdamSobieski) [Agenda+]

Explainer for the Decisions API

Reilly: the Decisions API answers questions with a menu of choices and returns a probability for each choice
… the Token Sushi demo compares rigid regular-expression matching, a capable but costly cloud LLM, and an on-device decision model
… high-confidence answers can be returned quickly, while uncertain cases are escalated to a larger model

[ Decisions API demo: Token Sushi ]

Semantic Embedder API

Semantic Embedder API

Reilly: the Semantic Embedder API supports similarity analysis by exposing numeric embedding vectors
… client-side embedding can reduce server load and avoid sending the original document to a server
… server-side search requires documents and queries to use the same model, so developers may need the exact browser model identity
… this raises update questions; developer feedback during the origin trial will inform the design

Ehsan: is the proposal ready for an early TAG review?

Reilly: yes; there are two opposing approaches in this space
… Semantic Embedder exposes embedding values but requires model interoperability, while Vector Store hides values in a local semantic-search database
… Vector Store cannot serve applications searching a much larger cloud dataset
… TAG feedback on the trade-offs between the two explainers would be valuable

Christian: TAG encourages early review of incubations to confirm that they are headed in the right direction

Reilly: Vector Store API explainer explainers-by-googlers/vector-store-api

Sushanth: perhaps the resolution on model calcification should be added to these two new explainers

Christian: +1

RESOLUTION: Request TAG feedback on Semantic Embedder API v Built-in Vector Store API.

Reilly: decision models read probabilities for candidate answers instead of generating text, making answers faster
… models focused on understanding and decisions can potentially be much smaller

Yunhao: what low-latency scenarios motivated the Decisions API?

Reilly: decision models can classify user intent, route support questions, or identify products related to a search query
… they can direct a request to a prepared site flow or a more expensive model, making on-device AI more practical

Sushanth: some decisions are category matching for a very small model, while others need more knowledge; can one model handle both?

Ehsan: does this move toward on-device personalization, and what exactly does a decision mean here?

Reilly: unlike prompting for one answer, a decision model returns a probability distribution over all supplied choices
… a high-confidence result can be used immediately; an ambiguous distribution can be escalated to a larger local or cloud model

Wrap up

Anssi: thank you for joining this Built-in AI APIs F2F meeting, it was a blast
… we'll share the minutes after a post-processing step and human review

Summary of resolutions

  1. Request TAG feedback on Semantic Embedder API v Built-in Vector Store API.
Minutes manually created (not a transcript), formatted by scribe.perl version 250 (Wed Oct 21 13:34:09 2026 UTC).