Browser AI trial details
This optional experiment uses a real small language model in a Web Worker on your device. It is not production-tested. It has no API key, hosted chat service, account access or external action tools.
Versions and download
Runtime: WebLLM 0.2.85, bundled and served with this site. Model: Qwen2.5-0.5B-Instruct, 4-bit MLC conversion. Model revision: 32ff081fe7e4dfe4ffb167b94c66fdf11e02b8ad.
The matching WebGPU model library is Qwen2-0.5B-Instruct-q4f16_1_cs1k from MLC’s v0_2_84/base build, pinned to binary repository revision 025bcaf3780fa8254f5e5efd3bfea0a5397248f4 and served by this site. The weights, tokenizer, configuration, manifest and model library total about 290 MB; the bundled runtime adds several more MB. The page rounds this to about 300 MB. Browser caching, overhead and implementation differences may change actual transfer and disk use.
Limits and privacy
Requests are text-only, capped at 600 characters. Replies are capped at 192 generated tokens with a 2,048-token context window. At most two recent exchanges are considered and older text is truncated. Closing or reloading the page loses the chat. AI output is displayed as plain text; the clickable source links are fixed, verified site pages.
The worker permits only GET downloads of the pinned model resources and the local model library. It disables other network transports, and blocks all fetch requests after model loading. There is no chat endpoint or cloud fallback. This is an implementation review and automated boundary test, not an independent security certification. The host and model download services still process technical requests, including IP addresses. See the Hugging Face privacy policy for its download service.
Stop & unload terminates the worker, ending loading or generation and releasing its model. Completed download files can remain cached. Remove downloaded model clears this trial’s matching files from site storage and verifies those entries are gone. It also clears chat and unloads the model. Your browser’s ordinary HTTP cache may retain downloads; its own browsing-data controls manage that separately.
Licenses and attribution
WebLLM © MLC LLM contributors and the Qwen2.5-0.5B-Instruct base model © Qwen contributors are licensed under Apache License 2.0. MLC’s conversion identifies Qwen’s base model; its repository does not supply a separate license file. This trial preserves the base model’s Apache notice.
- WebLLM Apache 2.0 license
- Qwen Apache 2.0 license
- loglevel MIT license
- Bundled runtime third-party notices
This experiment’s application integration, strict asset-download guard, bounded prompts and trial interface are TAIAIO modifications; the model weights and pinned WebLLM runtime are not fine-tuned or modified.