Chat, threads and history
How do I use chat threads and history?
It depends on your app build, operating system and model. One conversation by default, optional thread sidebar. Follow the steps below and check the current file record if a download fails.
Model weights are learned parameters; a runtime is the software that loads them. RAM is working memory, separate from disk space for downloaded files.
A checksum is a fingerprint of file bytes, not a safety guarantee. Read the verification policy. Run receipts, when present, are scoped to an exact file and setup.
Updated · Page-copy source revision 8871bab84e1cf80aea1686a944ec335078a6c036 · Source files: apps/web/lib/help-centre-content.mjs, apps/web/lib/help-centre-legacy.mjs, apps/web/lib/aeo-page.mjs · Full revision record
Key facts
- Desktop app
- Installed locally; each model file is a separate download with its own hash check.
- Checked label
- Means bytes matched a published SHA-256 on this computer; not a safety guarantee.
- Fit estimate
- Compares file size to memory; it is not a measured run or speed promise.
- Service status
- /status reports site probes; it does not diagnose your local runtime.
Evidence you can inspect (plain text)
- Inspect /verification for the site’s file-check policy.
- Open /artifact/hf-model-qwen-qwen3-8b/files for model exact filenames, availability and any recorded SHA-256.
- Open /evidence/artifacts/hf-model-qwen-qwen3-8b for run-readiness evidence tied to that record; absent receipts stay unmeasured.
- Open /api/artifacts/hf-model-qwen-qwen3-8b/card.md for machine-readable source facts (not a runtime test).
- Treat absent run receipts as unmeasured, not as proof of failure.
Unmeasured states are stated explicitly; this page does not invent run scores or compatibility guarantees.
Which date applies to which fact
- Committed page-copy sources:
- Upstream model metadata: Dates on artifact records when shown; not implied by page-copy revision
- Run receipts: Only on /evidence pages when recorded; otherwise unmeasured
Page-copy revision (Updated line above) identifies committed text sources, not upstream model edits or test outcomes.
Chats are stored on this computer. A model can give wrong answers, so check anything you rely on.
- Open Chat and choose a downloaded chat model. Wait for it to finish loading.
- Type in the box at the bottom. Press Enter to send. Use Shift and Enter for a new line.
- Ask a follow-up in the same box. Choose Stop to end a reply early.
- Use Copy on a reply, or Save this answer to keep it in your library.
Threads (optional)
Chat threads are off by default. In Settings, open About, find Experimental, and turn on Chat threads to show a sidebar per model, start New chat, and export a thread. Turning threads on reloads the app.