The Local LLM Leaderboard.

Every locally-runnable LLM, ranked by a transparent composite score across 34 benchmarked models.

Comparing two models? Use /compare/. Picking by hardware or task? See /best/. Cloud-vs-local cost? The cost calculator. Download the data: CSV · JSON (CC BY 4.0).

How this leaderboard is scored

Every model is rated 0–100 on a weighted composite: 45% quality (the MMLU benchmark), 25% efficiency (benchmark score per GB of VRAM — the metric that matters when the model runs on your hardware), 15% accessibility (does it fit 8/12/16/24 GB?), 10% license (can you use it commercially?) and 5% context window. Each sub-metric is min–max normalized across the 34 models tracked, so scores are relative to this set and recomputed on every update. VRAM figures and the efficiency metric assume a Q4_K_M quantization. Benchmark scores are sourced from each model’s published results (see its catalog page); models without the headline benchmark are listed as n/a and ranked last, never imputed. Full data: CSV / JSON under CC BY 4.0 — corrections welcome.

@misc{bestllmfor_local_llm_leaderboard_2026,
  author       = {Meguedmi, Mohamed},
  title        = {Local LLM Leaderboard 2026},
  howpublished = {\url{https://bestllmfor.com/leaderboard/}},
  year         = {2026},
  note         = {bestllmfor.com, CC BY 4.0}
}