Local models

Run AI writing entirely on your hardware. No API key, no data sent to Walkmark's servers or third-party cloud APIs.

Recording, editing, and export stay local either way. Only the Write action sends screenshots to a model — with local models that endpoint is on your machine or LAN.

Ollama

  1. Install Ollama and ensure the background service is running.
  2. In Walkmark, open Settings → Model and select On this Mac (Ollama provider).
  3. Download a vision model from the catalog (vision badge). Text-only models cannot read screenshots.
  4. Pick the model in the document toolbar and write as usual.

Walkmark talks to Ollama at http://127.0.0.1:11434 by default.

Catalog and downloads

The in-app catalog lists recommended vision models with size and RAM hints. Click Download to pull via Ollama; progress shows in Settings.

You can also enter a custom model id (e.g. my-model:tag) and pull or select it.

Pick based on available RAM:

Model (examples) Tradeoff
Moondream 2B Smallest; terse output
Qwen3-VL 4B / 8B Strong text reading for size
Gemma 3 / 4 Good non-English UI
Llama 3.2 Vision 11B Dense screen text

The catalog marks whether your machine likely has enough memory.

Concurrency

Ollama writes one step at a time regardless of the Concurrency slider — local models are slower and have less context headroom.

LM Studio and compatible servers

Any OpenAI-compatible HTTP API works:

  1. Start LM Studio (or vLLM, llama.cpp server, etc.) and enable the local server.
  2. In Walkmark Settings → Model, choose Custom endpoint.
  3. Set the base URL (e.g. http://127.0.0.1:1234/v1) and a vision-capable model id.
  4. API key is optional — leave blank if the server does not require one.

Walkmark sends chat-completions-style requests with image attachments the same way it does for cloud vision models.

Troubleshooting local models

Issue Fix
"Looking for Ollama…" Install and start Ollama
Model missing Download from catalog or ollama pull <name>
Out of memory Pick a smaller catalog entry
Garbled or empty steps Confirm the model supports vision; try Qwen3-VL or larger

See Troubleshooting.

Local vs cloud

Local (Ollama / compatible) Cloud (BYOK or Walkmark Cloud)
Data leaves your machine No (LAN only if remote server) Yes, when writing
Setup Install runtime + model API key or Cloud account
Speed GPU/CPU dependent Usually faster on frontier models
Cost Free after hardware Provider usage or subscription
Offline Yes, once model is downloaded No

Privacy

Local writing never sends frames to the public internet unless your custom endpoint is on a remote host you control. See Privacy and data.