Hugging Face's logo

Datasets:
localllmrig
/
localllmrig-open-data

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
1K - 10K
Tags:
local-llm
vram
gpu
llama.cpp
quantization
hardware
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
localllmrig-open-data / DATACARD.md
localllmrig's picture
localllmrig
Open dataset snapshot 2026-10-21 (CC BY 4.0)
c1013dc verified
preview code
|
Raw History Blame Contribute Delete
2.44 kB
metadata 
license: cc-by-4.0
pretty_name: localLLMrig open data
language:
  - en
tags:
  - local-llm
  - vram
  - gpu
  - hardware
  - llama.cpp
  - quantization
size_categories:
  - 1K<n<10K
configs:
  - config_name: fit_estimates
    data_files: data/fit-estimates.csv
  - config_name: hardware
    data_files: data/hardware.csv
  - config_name: models
    data_files: data/models.csv
  - config_name: gpu_price_per_gb_vram
    data_files: data/gpu-price-per-gb-vram.csv
  - config_name: showdown_outputs
    data_files: data/showdown/outputs-2026-10-17.csv
  - config_name: showdown_speeds
    data_files: data/showdown/speeds-2026-10-17.csv

Data card: localLLMrig open data (snapshot 2026-10-17)

Summary

Memory needs of 12 open-weight LLMs at four quantizations, 29 hardware options (GPUs, PC builds, Macs, laptops) with sourced EU prices, a 1,392-row model × hardware × quant fit and estimated-speed matrix, and a 60-output small-model showdown run on CPU.

Intended use

  • Planning which hardware can hold which model.
  • Comparing price per GB of VRAM with the price type in view.
  • Reading real small-model outputs side by side.

Not intended for

  • Treating estimated tok/s as benchmark results.
  • Price guarantees. Prices are dated snapshots.

Provenance

  • Models and hardware: localllmrig.com public API (/api/models, /api/hardware), fetched 2026-10-17.
  • Prices: pcprice.watch (used sold-listing medians), GPU PRIX and PriceSquirrel (EU new listings), Apple Store Belgium, ASUS eStore Belgium; URLs per row. Some rows are labelled estimates.
  • Showdown: run by localLLMrig on 17 Oct 2026; see METHOD.md §6.
  • Community measurements: third-party, quoted with source and caveats.

Fields with an estimated flag

need_gb_est, usable_gb_assumed, est_decode_tok_s, mem_*_total_gb_est, price_usd_fx_estimate, and PC build totals.

Known limitations

  • Single formula for decode speed; no prefill modelling; no runtime differences beyond a vendor utilization factor.
  • AMD utilization (0.40) is a conservative assumption.
  • Laptops with unpublished GPU-addressable memory are unscored.
  • Showdown speeds come from a shared CPU VM and vary per prompt.
  • Only 4 of 10 showdown prompts are checked; the rest are unscored.

License

CC BY 4.0. Model outputs remain subject to each model's license where applicable.

Maintainer

localLLMrig, https://localllmrig.com. Corrections via GitHub issues.