---
license: other
license_name: source-dependent
license_link: https://www.localscore.ai/about
tags: [benchmark, llm-inference, tokens-per-second, consumer-hardware]
---

# Local LLM speed benchmarks (community union)

Real measured local-LLM inference performance on consumer hardware:
(accelerator, model, quant, context) -> prompt tok/s, decode tok/s, TTFT.
Collected to train [FitCheck](https://huggingface.co/spaces/build-small-hackathon/FitCheck)'s
honest speed predictor ([methodology](https://arxiv.org/abs/2410.02425): LLM-Pilot,
leave-one-accelerator-out validation).

## Sources & attribution

- **LocalScore** (https://www.localscore.ai) — Mozilla Builders project by
  cjpais. The bulk of this dataset. Scraped politely from public pages; the
  upstream data license is unstated — if you are the maintainer and want this
  mirror changed or removed, open a discussion and it will be honoured fast.
- **llama.cpp community benchmarks** — GitHub discussions #4167 (Apple
  Silicon), #15013 (CUDA), #10879 (Vulkan/AMD/Intel), llama-bench tables.
- Hardware bandwidth specs joined from vendor product pages via the MIT
  spec tables of https://github.com/midudev/canirun.ai.

## Files

- `localscore_runs.jsonl` — run-level: hardware + model + averaged metrics.
- `localscore_tests.jsonl` — per-scenario rows (n_prompt 16-4096, n_gen
  16-3072) for context-dependence.
- `llamacpp_bench.jsonl` — fixed-workload community tables (validation set).
- `gpu_specs.json` — device -> memory bandwidth / VRAM (vendor specs).