Local Language Models for Voxt
Voxt keeps speech recognition and text enhancement separate. A local language model can take the transcript and clean filler, preserve terminology, rewrite tone, format notes, or turn spoken instructions into a useful prompt without sending that step to a remote provider.
The short version
- Qwen3.5 models cover lightweight, default, and higher-quality local rewrite tiers in the current evaluation notes.
- Gemma 4, GLM 4, and LFM2 provide non-Qwen alternatives for different language, size, and hardware preferences.
- Qwen3.6 27B and Qwen3 VL are high-memory or experimental choices rather than universal defaults.
- A local LLM still needs enough memory and may behave differently across quantization and runtime choices.
- For a simple voice workflow, choose the smallest model that reliably follows your cleanup or rewrite instruction.
Model map
| Model or family | Role | Useful when | Upstream source |
|---|---|---|---|
| Qwen3.5 2B, 4B OptiQ, and 9B OptiQ The main local rewrite tiers documented for lightweight, default, and higher-quality use. | Local LLM | Start with 4B for a balanced local workflow, then compare 2B or 9B against your own Mac and prompts. | Qwen models |
| Gemma 4 A non-Qwen local language-model path included as a fallback and comparison option. | Local LLM | Useful when you want to compare local rewrite behavior across model families. | Google Gemma |
| GLM 4 9B A Chinese and multilingual fallback choice in the current local model evaluation. | Local LLM | Consider it for Chinese-heavy prompts or when a non-Qwen model is preferable. | Z.ai GLM |
| LFM2 1.2B and 8B A1B Lightweight and mixture-of-experts local choices documented for model-size comparisons. | Local LLM | Use the smaller path for constrained hardware and test the larger option when quality matters more. | Liquid AI |
| Qwen3.6 27B and Qwen3 VL 4B Larger or visual model entries that should be treated as hardware-dependent and experimental in product copy. | High-memory or experimental | Reserve them for Macs with enough memory or experiments that actually need their additional capacity. | Qwen models |
Choose by task, not by parameter count
Voice cleanup is usually a short, constrained transformation. A smaller model may be a better everyday choice if it follows your instructions consistently and returns quickly. Larger models become more interesting when the task includes long context, complex formatting, or a more involved rewrite.
Keep your prompt and dictionary close
A model comparison is only useful when the instruction is held constant. Test the same personal dictionary, App Branch rule, product names, and output format across models so you are measuring the workflow rather than changing the prompt each time.
Local does not mean unlimited
Local processing avoids sending the model step to a remote provider, but it still consumes Mac memory, storage, CPU, or Apple Silicon GPU resources. Voxt's model picker and documentation should be treated as the current source of truth for what is available in a given release.
FAQ
What local LLMs can Voxt use?
The current Voxt local-model documentation includes Qwen3.5, Gemma 4, GLM 4, LFM2, Qwen3.6, and an experimental Qwen3 VL entry. Older or compatibility models are grouped in the client documentation rather than presented as the primary recommended catalog.
What is a good local LLM for voice rewriting on Mac?
Qwen3.5 4B is the documented balanced starting point, but the right choice depends on your Mac memory, language, prompt complexity, and tolerance for latency. Compare it with the 2B or 9B tier using the same real voice samples before deciding.
Can Voxt use an LLM without sending voice data to the cloud?
Yes, when both the selected ASR channel and the enhancement channel are local and the workflow does not call a remote provider. Check Voxt's privacy and data-flow documentation because different channels can have different boundaries.