|
Download eval/local_intelligence/model_candidates.md from arach/training-lab: direct link, hf CLI and curl.
- Browser
- Download file 674 Bytes
-
https://huggingface.co/arach/training-lab/resolve/main/eval/local_intelligence/model_candidates.md
- Command line
-
hf download hf://arach/training-lab/eval/local_intelligence/model_candidates.md
-
curl -L -o model_candidates.md https://huggingface.co/arach/training-lab/resolve/main/eval/local_intelligence/model_candidates.md
674 Bytes
Recommended Open Models
Current shortlist for the Talkie local-intelligence eval pack on this machine:
mlx-community/NVIDIA-Nemotron-3-Nano-4B-4bit- Plain
mlx-lmmodel - Good fit for structured JSON tasks without pushing memory too hard
- Plain
mlx-community/gemma-4-e4b-it-mxfp8- Runs through
mlx-vlm - Stronger modern general-purpose open model, but heavier than the Nemotron 4B path
- Runs through
mlx-community/Qwen3.5-2B-4bit- Existing local baseline already partly in cache
- Useful as the control model for before/after comparisons
Suggested first pass:
Nemotron 4Bon all 24 cardsGemma 4 E4Bon all 24 cardsQwen3.5 2Bas baseline