[FEATURE] WebLLM model provider (on-device inference via WebGPU) #2481

Problem Statement

The Strands TypeScript SDK does not ship a first-class model provider for WebLLM. WebLLM runs quantized LLMs entirely in the browser via WebGPU, with model weights cached in IndexedDB/CacheStorage after the first download. Without a dedicated provider, users who want on-device, offline-capable agents have to wire WebLLM up themselves.

Use Case

• Building browser-based agent apps that run fully offline after the initial model download
• Privacy-sensitive deployments where prompts/responses must not leave the user's device
• Eliminating per-call API costs by running inference on user hardware
• Demo/educational apps that work without any cloud credentials

Proposed Solution

A WebLLMModel provider under @strands-agents/sdk/models/webllm that:

• Wraps @mlc-ai/web-llm (MLCEngine / CreateMLCEngine)
• Implements the standard Model interface (streaming, tool use, stop reasons, usage)
• Exposes helpers for cache management: downloadWebLLMModel, isWebLLMModelCached, deleteWebLLMModel, listWebLLMModels
• Surfaces InitProgressReport via an onProgress callback so apps can render download progress
• Guards against non-browser environments (WebLLM requires WebGPU)

Workaround

Today you can use VercelModel with the community webllm-ai-provider package (Vercel AI SDK v2 provider spec):

import { Agent } from '@strands-agents/sdk'
import { VercelModel } from '@strands-agents/sdk/models/vercel'
import { webllm } from 'webllm-ai-provider'

const agent = new Agent({
  model: new VercelModel({ model: webllm() }),
})

This works but has caveats:

• webllm-ai-provider is a third-party package at 0.0.1 with a single maintainer and ~2 weekly downloads
• No built-in cache management API (download-before-use, evict, list-cached)
• Progress reporting goes through the provider rather than a first-class onProgress hook
• WebGPU/browser-only environment checks are the user's responsibility

Alternatives

1. Keep recommending the VercelModel + webllm-ai-provider workaround in docs only
2. Consume @mlc-ai/web-llm directly via a custom Model subclass in user code
3. Ship a dedicated WebLLMModel provider (proposed)

Additional Context

• WebLLM repo: https://github.com/mlc-ai/web-llm
• webllm-ai-provider on npm: https://www.npmjs.com/package/webllm-ai-provider
• Default models listed in WebLLM's prebuiltAppConfig