Speech-to-text is only half the pipeline — the second half rewrites the transcript according to your workspace. That step can also run locally, which is what makes a fully offline setup possible.
What is offered
Nine curated models, all quantised GGUF builds. A representative selection:
| Model | Size | Suits |
|---|---|---|
| Phi-4 Mini 3.8B | 2.5 GB | Recommended starting point |
| Llama 3.2 3B | 2.0 GB | Modest machines |
| Qwen 3.5 4B | 3.1 GB | A step up in quality |
| Llama 3.1 8B | 4.9 GB | 16 GB RAM or more |
| Phi-4 14B | 9.7 GB | 16 GB RAM or more |
| Gemma 4 26B | 17.0 GB | 32 GB RAM or more |
Sizing it
The model has to fit alongside your speech model and everything else you are running. As a rough guide, allow the file size plus a couple of gigabytes for context. The context window defaults to 16,384 tokens and is adjustable on the same page — lowering it reduces memory use.
The runtime offloads to GPU where it can and falls back to CPU automatically. Local refinement adds latency to every dictation, so watch the total against the sub-second feel you are used to.
When it is not ready
If local mode is selected but no model has finished loading, InkSpoke does not fail — it injects the raw transcription instead. This is deliberate, but it does mean that "my tone settings are being ignored" can really mean "the local model was not loaded".