Local models

Wrengle Local uses GGUF local agent models managed from Settings → Models → Local agent models. The local model path is separate from bring-your-own-key OpenAI/Anthropic settings and from cloud transcription keys.

Voice analysis and the built-in assistant default to Auto, so a local agent model is not required when a connected cloud AI key is confirmed available in the OS keychain. Auto prefers the most recently connected or successfully verified AI provider, then OpenAI → Anthropic, and uses Wrengle Local only when no eligible cloud route is available. Selecting Wrengle Local explicitly remains sticky and still requires a downloaded model; Reset returns the surface to Auto.

Dictation speech has its own Auto route and local Whisper downloads. It prefers the most recently connected or successfully verified speech provider, then OpenAI → Deepgram → ElevenLabs, before Local Whisper. These automatic voice and assistant defaults do not change meetings: live meeting captions and the default final transcription remain local-first, cloud meeting final-pass transcription stays opt-in, and meeting reports use their own configured model.

Model catalog

Wrengle ships a Qwen3-family model catalog:

The app checks system memory and recommends the largest catalog model your machine is expected to handle. Machines below the Tiny threshold cannot use the local agent path and can use Auto with a connected cloud provider instead.

Download and storage

Model downloads come from Hugging Face catalog URLs. Wrengle stages each download as a temporary .part file, verifies the SHA-256 checksum, and only then promotes it to the final GGUF path. A checksum mismatch deletes the staged file instead of installing an unverified model.

Local agent models are stored under the app data model directory in models/llm/, separate from Whisper and embedding assets. The app checks free disk space before downloading and requires the model size plus extra headroom.

Managing models

From Settings → Models → Local agent models you can:

• download a catalog model;
• choose the active downloaded model;
• remove downloaded models to free disk space;
• remove orphaned GGUF files that are no longer part of the catalog.

Setting a model active validates that its on-disk path exists before committing the selection.

Privacy

Local model inference keeps prompts and context on your machine when the selected workflow actually runs on Wrengle Local. Auto is not a promise of local execution: with an eligible key it resolves to a cloud provider, which receives the request data documented for that workflow. Downloading a model contacts the catalog download host, and provider-backed workflows still follow the provider boundary documented in AI and privacy.