Spaces:
Running
Running
| title: Local Model Explorer + quant picks | |
| emoji: 🎯 | |
| colorFrom: indigo | |
| colorTo: blue | |
| sdk: static | |
| pinned: false | |
| tags: | |
| - gguf | |
| - llama-cpp | |
| - local-llm | |
| # Local Model Explorer + quant picks | |
| Same idea as [codelion's Local Model Explorer](https://huggingface.co/spaces/LocalLLaMA/local-model-explorer): | |
| pick your hardware, get the models that fit. This clone goes one further — | |
| **every model also gets its exact best quant**, ranked by measured quality | |
| instead of size alone. | |
| Fit + catalog data load live from the original explorer's API. Quant ranking | |
| comes from `quant_scores.json`: KL-divergence vs full precision over 150 | |
| GGUFs (LFM2.5-1.2B, Qwen3.5-0.8B, gemma-3-1b-it). KV precision comes from | |
| `kv_scores.json`: 12 K/V pairs benched on 3 models (KL vs f16 @4k, PPL + | |
| VRAM @64k), wrapped client-side around the upstream f16-only API with | |
| q5_0/q4_1 preselected. Method in [quant-bench](https://github.com/) | |
| `SUMMARY.md` / `PRIORITY.md`. | |
| Rules: single auto pick per model — best measured quality fully resident | |
| on your hardware, no band caps (if Q8_0 fits, it wins). Dominated quants | |
| (`Q4_0`, `Q4_1`, …) lose to their better twin when it's in the same repo. | |
| Logic in `rankQuants.js` (mirrors `rank_quants.py`). | |