Meeting schedule in Greenwich Mean Time (GMT). Remote participants, please confirm your local start time using the timezone table below.

Working Group home | charter

WebNN API spec | explainer | open issues | implementation report

▶️ 09:00 (GMT) Welcome & direction

Intros & agenda bashing

Confirm the agenda and identify any attendance constraints.

Charter orientation

Current baseline, major milestones, expected changes.

• WebNN API 22 January 2026 CR Snapshot and Chrome/Edge Developer Trials
• API stabilizing, key breaking changes/new features since Kobe: minimum data types and rank ranges PR #910, initial WebGPU interop PR #942, and security limits PR #926 / PR #933
• Status of Kobe resolutions: dynamic shapes #883, subgraphs #907, streaming constants #901, and compute policy #911
• Charter next steps

Proposed breaking changes

Confirm the compatibility plan, owners.

• Unblock MLComputePolicy PR #923
• Land the rank relaxation and operator renames from #904
• Decide whether to remove constant tensor support #940
• Discuss any blockers for the resolution allowing zero-size dimensions #391

Device selection & effective compute policy

Converge on a proposal to develop, document considered alternatives.

• Compare the four #934 proposals: compilation/runtime metrics, latency vs throughput, strict hints, and low-latency / precision hints
• Decide how intendedComputePolicies, performance reporting, and the DAOP incubation fit together
• Validate against use-case hints #902 and real-time workloads #898
• Address privacy and fingerprinting concerns around effective policy or device exposure
• Related breakout: Giving the user control over where AI models run in the browser w3c/tpac2026-breakouts#72

⏸️ 10:30 (GMT) Break

Drinks and snacks: Pre-function + Bbar, Ground Floor.

▶️ 11:00 (GMT) LLM enablement

Dynamic shapes

Discuss dynamic shapes implementation experience and any further spec refinements.

• Review #883 and dynamic-shape explainer PR #945 as appropriate
• Triage PR #945 questions on tensor-derived sizes, fingerprinting, and symbolic shape inference
• Compare Core ML and LiteRT implementability and prototype experience
• Define the security review needed for dispatch-time validation

Subgraphs & operator set

Define the minimum subgraph API experiment, JS examples, and performance evidence.

• Review the #907 proposal and Grouped Query Attention example
• Core operator set #573
• Consider a loop operation for state-space models #559
• Agree to not remove operators #689 before subgraphs exist and are measured

Low precision & quantization

Discuss new data types and quantization interoperability.

• Low-precision data types #930, explainer PR #938, and spec PR #956 next steps
• Agree criteria for adding data types and how unsupported types are reported
• Decide next steps for optional zeroPoint #943, Core ML constant-only constraints #822, and blockwise expansion #903

New proposals

Introduce new proposals to fill in the gaps in LLM enablement.

• Build-time metadata #965
• Stateful model #966
• Related implementation experience: a dynamic language model (Qwen3-4B) on the Intel NPU via WebNN

⏸️ 12:30 (GMT) Lunch

Sussex restaurant + Bbar, Ground Floor.

▶️ 13:45 (GMT) Implementations & ecosystem

WebNN implementation status by: operations | frameworks | backends Lite RT, Windows ML, Core ML | hardware accelerators | architecture

Demos & implementation status

Record interoperability blockers, implementation experience to inform specification priorities.

• Chrome/Edge Developer Trial/prototyping feedback #964
• RustNN running BabyLM models on constrained devices
• More demos: LiteRT.js integration, NPU stem separator, Transformers.js demos
• Backend status: LiteRT, Windows ML, and Core ML

Customer stories

Select the highest-impact framework or model gaps to feed into the WebNN roadmap.

See WebNN Awesome for community feedback, demos, presentations, samples, testimonials, tutorials, videos, and more.

• onnx2webnn: top-50 Transformers.js model validation and operator conversion

Weights & memory

Discuss prototyping interest for streaming and caching, with agreement on measurements e.g. peak memory and startup time.

• Streaming constants and separation of graph build from weight loading #901
• MLGraph caching #807
• Relationship to hash-addressed Cross-Origin Storage, without expanding WebNN's scope into a storage API

WebGPU interop v2

Agree the v2 API direction for device matching and usages, and identify any concrete cross-group coordination requests.

• interopDevice for zero-copy selection #941
• Exported GPUBuffer usage flags #954
• Timelines #529
• Sample code #343
• Coordination with the GPU for the Web WG

⏸️ 15:00 (GMT) Break

Drinks and snacks: Pre-function + Bbar, Ground Floor.

▶️ 15:30 (GMT) Security, Conformance & wrap-up

WebNN implementation report | WPT test suite | webnn report

Security hardening

Confirm normative algorithms and owners for the remaining security-related spec and WPT changes.

• Relational limits for convolution and pooling #928
• Refresh Security Considerations and out-of-bounds handling #927
• Discuss full-sequence descriptor validation #949 / PR #955 as appropriate.

Numerics & WPT

Agree the normative-vs-test tolerance split and spec and WPT changes.

• Accumulator precision explainer #948
• Trigonometric accuracy #914
• Stable decompositions #947 / PR #957
• WPT tolerances #950 and boundary classification #960
• GPU-backed WPT runners update (26 March resolution)

Issue triage slate

Check owners, interest, close or explicitly defer.

• Resolved but no spec PR: #935, #919 (#391/#904/#928 are covered in dedicated slots)
• LiteRT backend gaps: #876, #877, #878
• Close candidates: #853, #879, #124, #900
• Older interop backlog: #180, #444, #481, #486, #487, #691

WebNN horizontals

Identify required horizontal reviews, owners, and supporting WebNN text or evidence.

• Privacy and fingerprinting risks from compute-policy, performance, and opSupportLimits exposure
• Sustainability impact follow-up #861

Wrap-up

Confirm the resolution summary, next CR Snapshot target.

• Synthesis of resolutions, actions, known risks
• Next Candidate Recommendation Snapshot target

To be triaged before F2F

Resolve asynchronously or bring to F2F agenda.

Possible additional topics: batched gemm #929, inferred convolution groups #959, mixed-type pow #939, and custom operations #6.

⏹️ 16:45 (GMT) End

• 🍽️ Group dinner (venue TBA)

Meeting schedule in Greenwich Mean Time (GMT). Remote participants, please confirm your local start time using the timezone table below.

▶️ 08:30 (GMT) Start

Intros & agenda bashing

Confirm participants, scribes, agenda, and decision-making process.

WebMCP since Kobe

Establish a shared baseline for the day.

• Celebrate major spec progress since Kobe
• Acknowledge wide review, standards positions, implementation momentum
• Standardization path and the Agentic Web WG Charter draft

API shape & multi-stakeholder support

Converge on the API direction and key design considerations.

• Generic communication and Tool-object direction #236
• API design questions: headless use, origin isolation, progress, and schema format #237

Tool discovery & cross-origin exposure

Agree the discovery and exposure model to develop.

• Discovery beyond one traversable navigable #227
• Host-controlled iframe access #317 and cross-origin agents #52
• Security and Privacy: Two-sided opt-in PR #326

Declarative API next steps

Decide the home and maintenance plan for declarative WebMCP work.

• Spec PR #338, native support concerns #144, and JS-driven forms #138

⏸️ 10:00 (GMT) Break

Drinks and snacks: Pre-function + Bbar, Ground Floor.

▶️ 10:30 (GMT)

Developer feedback & ecosystem

Extract actionable implementation and developer feedback.

• Origin Trial feedback
• WebMCP Challenge takeaways from the 5000+ submissions
• Current implementation status
• WebMCP polyfill

Cross-document continuations & page context

Align navigation continuity, tool stability, and page-context updates.

• Converge continuation designs: tokens PR #327 vs continueAfterNavigation #135
• Tool stability #328
• Agent synchronization with page context #312 and context loss on tool changes #262

Tool contracts: schemas, results, annotations

Agree on tool schema, result, and annotation semantics.

• outputSchema PR #254 and #9 closure
• Browser validation and the JSON Schema subset #92
• High-risk reads #331
• Input length limits #73, text units #219 and PR #265

Human-in-the-loop & user activation

Choose a viable path and scope for user interaction.

• Elicitation #165 and PR #204
• User activation #62 and human approval bypass #288

⏸️ 12:00 (GMT) Lunch

Sussex restaurant + Bbar, Ground Floor.

▶️ 13:15 (GMT) Horizontal groups and coordination (joint sessions)

Privacy & security

Joint with Privacy WG / Privacy IG / Security IG.

Agree concrete WebMCP mitigations, external guidance, and review follow-ups.

• Privacy mitigations #313 and ecosystem-risk guidance PR #334
• Security Threat Model for WebMCP, related discussion #154, prompt injection #11

WebMCP and WebExtensions

Joint with WebExtensions CG.

Discuss integration use cases, boundaries, and next design steps.

• Separate TPAC breakout: Integrating WebMCP into WebExtensions
• Related: extension enumeration and invocation #74, external MCP client and permissions considerations #23

Wide review follow-ups

Record remaining review requests, responses, and target dates.

• TAG design-reviews#1238
• Accessibility w3c/a11y-request#189 and related discussion #65
• Internationalization w3c/i18n-request#338

▶️ 14:00 (GMT)

Built-in AI APIs overview

• Implementation status
• Origin/Dev Trials and developer feedback
• Standards positions and wide review

Prompt API

Spec | Explainer | Issues | PRs

• Model selection and availability #169
• Message format for tool calling, thinking, and multimodality #209
• Sampling modes #203, thinking effort #222, and fastest mode #219

Writing Assistance APIs

Spec | Explainer | Issues | PRs

• Privacy of a paid cloud option #84
• Performance preference limits and cross-API alignment #96 / PR #103
• contextUsage and contextWindow alignment with Prompt API #109

Translator and Language Detector APIs

Spec | Explainer | Issues | PRs

• Input formats for translation and language detection #60, #82, #83
• Definition of on-device models #72
• Batched translation #85

Proofreader API

Spec | Explainer | Issues | PRs

• Multiple correction options #35 and native spellchecker fallback #36
• HTML input #32 and custom dictionaries #20

🌱 Proposals for new incubations

• Review and discuss any new proposals with traction:
  • Decisions API #20
  • Semantic Embedder API explainer

🫶 Wrap up

• Synthesis of resolutions, next steps

⏹️ 15:00 (GMT) End

You may also be intested in:

15:30 - 16:30 (GMT)

Breakout sessions

🤝 18:00 (GMT)

Developer Meetup (registration)