A curated catalogue of community 4-bit conversions, downloaded straight from Hugging Face and verified on arrival. Pick by memory, by use case, or let the recommended default do the work.
The recommended default. A 4-bit reasoning model at 2.4GB, quick enough to feel instant on Apple Silicon.
| Model | Size | Suited to |
|---|---|---|
| NVIDIA Nemotron Nano 4B | 2.4GB | Recommended, 16GB |
| NVIDIA Nemotron 3 Nano 30B A3B | 16.6GB | Use with caution, needs 48GB |
| Llama 3.1 8B Instruct | 4.2GB | Recommended, 16GB |
| Llama 3.1 Nemotron 70B Instruct | 37.0GB | Use with caution, needs 64GB |
| Qwen2.5 7B Instruct | 4.0GB | Recommended, 16GB |
| Qwen3 8B | 4.3GB | Recommended, 16GB |
| Qwen3 14B | 7.8GB | Use with caution, needs 32GB |
| Qwen3 Embedding 0.6B | 335MB | Required for document analysis |
| Gemma 3 12B Instruct | 7.5GB | Use with caution, needs 32GB |
| Phi-4 mini instruct | 2.0GB | Recommended, 16GB |
| DeepSeek R1 Distill Qwen 7B | 4.0GB | Recommended, 16GB |
| Mistral 7B Instruct v0.3 | 3.8GB | Recommended, 16GB |
Every model is a 4-bit conversion, sizes shown as downloaded. Guidance, not gatekeeping: the in-app catalogue reads your machine's memory and tells you what will run comfortably before you download anything.
Every model in the catalogue is an open source model in a published 4-bit conversion. Downloads are resumable, so a dropped connection costs you minutes, not gigabytes.
Each downloaded file is integrity-checked against a per-file SHA-256 hash before it can ever be loaded. If a byte is wrong, the model does not run.
Pathway Lite runs community 4-bit conversions, mostly from the MLX community on Hugging Face. We do not retrain or fine-tune models. We select published conversions, verify them, and run them as they are.
Rewrite the system prompt to shape tone, format and behaviour, per chat or per project. No vendor defaults you cannot see or change.
Turn it down for precise, repeatable answers, or up for looser, more creative output. Adjust it mid-conversation and see the difference immediately.
Occasional notes on new models, new skills, and new releases.
Llama 3.1 8B Instruct
Qwen2.5 7B Instruct
Gemma 3 12B Instruct
Phi-4 mini instruct
DeepSeek R1 Distill Qwen 7B
Mistral 7B Instruct v0.3