Local AI Scope: local LLM memory fit, file sizes and language token cost
Open data from Local AI Scope — https://localaiscope.com/data/

Version:              2026-10-23.2  (a version is a date, .2 if the data changed twice that day; it only changes when the data does)
First published:      2026-10-08
Models measured:      2026-09-30 to 2026-10-23  (each model's date: measured_on in models.csv)
Machine data updated: 2026-10-23
Dataset file built:   2026-10-23T08:17:57Z
Contents:             26 models x 18 machines = 468 pairs
Method:               https://localaiscope.com/en/methodology/  (how every figure is obtained, and its evidence label)

FILES
  dataset.json  (68499 bytes)
    The complete dataset: byte for byte the file the Model Finder and the model and machine sheets are computed from (the Finder receives the same data embedded in its page).
    sha256 254a80302521ce4fa38124dff827110fb335eca677e6266fae98aaeb19e4df51
  models.csv  (9566 bytes, 26 rows)
    One row per model: published GGUF file size per quantization (GiB), maximum context, attention layout, license and primary source.
    sha256 1d1bdf4c004774bc6e3c1717b01d7118de9faf40983a6e46660abc230396c181
  machines.csv  (7523 bytes, 18 rows)
    One row per machine profile: memory, system reserve, memory usable by the model, bandwidth and compute with their evidence label and source, manufacturer price where published.
    sha256 38ac34dc55929098560d568d5b5e0afeefd99f9131ebf4f8112445c71050a0ba
  fit.csv  (37776 bytes, 468 rows)
    Every machine x model pair: does it fit at the reference quantization, up to which context, and the lighter quantization that fits when the reference does not.
    sha256 a4de61a89392ba17d4798643095d2dfbbdc422494d37191055734218b588e72f
  language_cost.csv  (18632 bytes, 520 rows)
    Measured token cost of saying the same text in ES, FR and DE relative to EN, per model and type of text, with the model's own tokenizer.
    sha256 553ac60dcbc9e671f26ddc9c9a1baeabebf539cc1dfc649afd17d337700ead90

  SHA256SUMS.txt  checksums of the files above:  sha256sum -c SHA256SUMS.txt
  CITATION.cff    the citation below, machine-readable
  manifest.json   all of this, as JSON

HOW TO READ THE COLUMNS
  GiB = 2^30 bytes. File sizes are those of the GGUF files actually published for each quantization
  (gguf_repo). 'fits' uses the reference quantization (Q4_K_M, or the nearest one the author publishes)
  and the smallest context step (4,096 tokens): file + context memory + runtime overhead <= usable memory.
  Context steps are 4k, 8k, 16k, 32k, 64k, 128k, 256k, 512k and 1M tokens, capped by the model's maximum;
  an empty fits_* cell means the model does not accept that context at all.
  usable_memory_gib = memory - system reserve, and never more than the manufacturer's GPU cap if there is one.
  Context memory = 4 bytes x kv_heads x head_dim per layer and token (fp16 key + value); sliding layers stop
  at sliding_window_tokens. When kv_bytes_per_token_full/_sliding are filled (Gemma 4: other head size in the
  global layers, layers that share another layer's cache), use them instead: bytes per token of all those layers.
  layers_linear_attention keep a fixed-size state (linear attention, Mamba-2) that does not grow with context
  and is not counted; layers_no_cache (mixture-of-experts or MLP blocks of hybrid models) keep nothing.
  tokens_vs_en = how many tokens it costs to say the same text in that language, with English = 1.
  tokens_per_word is per word of that language and is NOT comparable across languages; use tokens_vs_en.
  Evidence labels follow the methodology page: Vendor declared, Derived, Estimated, Measured by Local AI Scope.
  Speeds are not included: on the site they are labelled Estimated and never decide whether a model fits.

LICENSE
  The data files in this folder are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0):
  https://creativecommons.org/licenses/by/4.0/
  You may copy, adapt and reuse them, including commercially, crediting "Local AI Scope, localaiscope.com" with a link,
  linking to the license and indicating any changes.
  The license covers these data files only, not the text, design, code or photos of the website.
  Manufacturer figures are facts, each with its source in its row. Each model's own license (column
  'license') belongs to the model's author and governs the model, not these files.

HOW TO CITE
  Local AI Scope (2026). Local AI Scope: local LLM memory fit, file sizes and language token cost (version 2026-10-23.2) [Data set]. localaiscope.com. https://localaiscope.com/data/ Licensed under CC BY 4.0.

  @misc{localaiscope_20261023_2,
    author       = {{Local AI Scope}},
    title        = {Local AI Scope: local LLM memory fit, file sizes and language token cost},
    year         = {2026},
    version      = {2026-10-23.2},
    howpublished = {\url{https://localaiscope.com/data/}},
    note         = {Data set, CC BY 4.0}
  }

CORRECTIONS
  If you reproduce a figure and get something else, write to the address in the legal notice:
  https://localaiscope.com/en/legal-notice/  — corrections are published with their date in the change log.