Datasets:

localllmrig
/

localllmrig-open-data

license: cc-by-4.0
pretty_name: 'localLLMrig open data: local LLM memory needs, hardware and EU GPU prices'
language:
  - en
tags:
  - local-llm
  - vram
  - gpu
  - llama.cpp
  - quantization
  - hardware
  - benchmark
  - prices
size_categories:
  - 1K<n<10K
configs:
  - config_name: fit_estimates
    data_files: data/fit-estimates.csv
  - config_name: models
    data_files: data/models.csv
  - config_name: hardware
    data_files: data/hardware.csv
  - config_name: gpu_price_per_gb_vram
    data_files: data/gpu-price-per-gb-vram.csv

localLLMrig open data: local LLM memory needs, hardware, EU GPU prices and a small-model showdown

Open, citable data behind localllmrig.com: how much memory popular open-weight LLMs need, which GPUs and Macs they fit on, estimated decode speed, EU GPU prices with sources, and real outputs from six small open models on ten prompts.

License: CC BY 4.0. Credit line: Data: localLLMrig (https://localllmrig.com), CC BY 4.0.

Cite as: localLLMrig open data, CC BY 4.0 — https://localllmrig.com/open-data?utm_source=huggingface

Live site exports (`data/site-exports/`)

The same files the site publishes at localllmrig.com/open-data, copied as of 21 Oct 2026:

File	What it is
`fit.csv`	Which open-weight model fits on which machine, at each quant (estimate)
`gpu-price-per-gb-vram.csv`	Graphics card prices per GB of VRAM, used and new, plus Macs (also at /data/gpu-price-per-gb-vram)
`hardware.csv` / `hardware.json`	Machines and graphics cards with price, date and source
`models.csv` / `models.json`	Open-weight models and their sizes
`showdown/data.json`	Free LLM Showdown: every prompt and answer
`showdown/speeds.csv`	Free LLM Showdown: speeds and scores

How it's updated

The site is the source of truth. When prices or models change on localllmrig.com, the files at /open-data/* change with them.
This repository takes a dated snapshot of those files. Each snapshot is one commit, with the date in the commit message; price_as_of / date columns give the date per row.
Prices move. Always check the source_url of a row before buying anything.
To pull the latest files yourself: curl -O https://localllmrig.com/open-data/fit.csv (same for the other file names).

Estimates are not measurements. Every memory figure and every est_decode_tok_s value comes from a published formula (see METHOD.md). The only measured speeds are the showdown CPU numbers (one run, CPU-only cloud VM, not representative of GPU speed) and the third-party community rows, which are labelled with their source.

Files

File	Rows	What it is
`data/models.csv`	12	Model architecture (layers, KV heads, head dim, max context) and estimated memory need at 8,192 tokens for Q4_K_M, Q5, Q8 and FP16.
`data/hardware.csv`	29	GPUs, PC builds, Macs and laptops: memory, bandwidth (with source where stated), price in EUR incl. VAT with label, date and source URL. USD is an FX estimate.
`data/gpu-price-per-gb-vram.csv`	7	Single cards ranked by EUR per GB of VRAM. Used medians and new in-stock lows are labelled; do not compare them as if they were the same thing.
`data/fit-estimates.csv`	1,392	Every model × hardware × quant: estimated need, assumed usable memory, fit (`full`, `offload`, `no`, `unscored`) and estimated decode tok/s.
`data/community-measurements.csv`	3	Third-party measurements quoted on the site, with source and caveats. Not localLLMrig measurements.
`data/showdown/`	6 models × 10 prompts	Free LLM Showdown (17 Oct 2026): prompts, full unedited outputs, hand checks on 4 objective prompts, CPU speeds, GGUF files with sha256.
`data/localllmrig-data.json`	—	Everything above in one JSON file, plus the formula constants.
`snapshot/`	—	Raw inputs: `/api/models` and `/api/hardware` as fetched on 2026-10-17, and the showdown JSON.
`scripts/build.py`	—	Rebuilds `data/` from `snapshot/` (or `--fetch` from the live API). Python 3 standard library only.
`scripts/validate.py`	—	Cross-checks the fit/speed matrix against the live `/api/recommend` endpoint.

Quick start

python3 scripts/build.py          # rebuild from the snapshot
python3 scripts/build.py --fetch  # refresh from https://localllmrig.com/api/*
python3 scripts/validate.py       # compare with the live recommender

import pandas as pd
fit = pd.read_csv("data/fit-estimates.csv")
fit.query("model_id == 'qwen2.5-32b' and quant == 'Q4' and fit == 'full'")

Snapshot

Prices: as of 14–17 Oct 2026 (per row price_as_of). Prices move; check the source URL before buying anything.
Formula version: localLLMrig lib/formula.ts, priceAsOf 2026-10-14.
Showdown run: 17 Oct 2026, llama.cpp b11707, Q4_K_M GGUFs, temperature 0.

Cite

Archived on Zenodo: version DOI 10.5281/zenodo.23535233 (snapshot 2026-10-21); concept DOI 10.5281/zenodo.23535232 (always resolves to the latest version).

localLLMrig (2026). localLLMrig open data: local LLM memory, hardware and EU GPU prices (snapshot 2026-10-17). https://localllmrig.com. CC BY 4.0. https://doi.org/10.5281/zenodo.23535232

Licensing notes

The compilation, formula outputs and documentation are CC BY 4.0.
Showdown answers are outputs of third-party models. Each model's own license applies to use of that model (see license and license_url in data/showdown/speeds-2026-10-17.csv). Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Gemma is provided under the Gemma Terms of Use (ai.google.dev/gemma/terms).
Prices are facts recorded from the named public sources, with links. Spec figures link to vendor pages where available.

Corrections

Open an issue with the row, what is wrong, and a source link. Corrections are applied to the site and the next snapshot.