Open data

The human page is mlx.app/data. The files are static: built once per deploy from the same catalogs and the same estimate engine as the pages, served from the CDN, and never edited by hand.

#GET /data/index.json

The manifest: what's here, its version, licence, every dataset's columns and each file's URL, size and SHA-256. Note the path: the datasets live under mlx.app/data, not under /api.

curl https://mlx.app/data/index.json
curl -O https://mlx.app/data/macs.csv && shasum -a 256 macs.csv

const m = await (await fetch("https://mlx.app/data/index.json")).json();
for (const d of m.datasets) for (const f of d.files) console.log(d.id, f.format, f.bytes, f.sha256);

import hashlib, requests

m = requests.get("https://mlx.app/data/index.json").json()
for d in m["datasets"]:
    for f in d["files"]:
        body = requests.get(f["url"]).content
        assert hashlib.sha256(body).hexdigest() == f["sha256"], f["url"]
print("version", m["version"], "checksums match")

{
  "name": "MLX open data: Apple Silicon Macs, open AI models and what runs where",
  "url": "https://mlx.app/data",
  "version": "2026-10-19",
  "dateModified": "2026-10-19",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "attribution": "MLX Network, https://mlx.app/data",
  "method": "https://mlx.app/compare#cp-how",
  "datasets": [
    {
      "id": "macs",
      "name": "Apple Silicon Mac configurations",
      "description": "Every Apple Silicon chip from M1 to M6 in every memory size Apple sells it with: unified memory, memory bandwidth, GPU core options and the Macs it ships in, read from Apple's Tech Specs pages, with the source of every row.",
      "dateModified": "2026-10-19",
      "rows": 58,
      "columns": [
        { "name": "mac", "type": "string", "description": "Identifier: chip and memory size, as in mlx.app/mac/<mac>" },
        { "name": "chip", "type": "string", "description": "Chip name" }
      ],
      "files": [
        { "format": "json", "url": "https://mlx.app/data/macs.json", "bytes": 45839, "sha256": "76320724a7fac7650f78f2c51a08b5f74e9a5d6791d6cfd51a6a3795e3ff81f2", "mime": "application/json" },
        { "format": "csv", "url": "https://mlx.app/data/macs.csv", "bytes": 26453, "sha256": "e8a976f183d2a11401e038cbb4b2c6fefa9200f2a0039b4cc771b5893eade384", "mime": "text/csv" }
      ]
    }
  ]
}

datasets is trimmed to one dataset and its columns to two here; the live manifest lists all three datasets with every column and its description.

#The files

Each URL always serves the newest version. Every response carries:

• X-Dataset-Version: the version, the same as version in the manifest.
• Link: the licence (rel="license") and mlx.app/data (rel="describedby").
• Content-Disposition with a file name that includes the version (mlx-macs-2026-10-19.csv), so a saved copy says which one it is.
• Access-Control-Allow-Origin: *.

#JSON

A JSON file is one object: name, id, version, dateModified, url, license, attribution, description, method, sources (each { text, href }), columns (as in the manifest) and rows, an array of objects keyed by column name. List columns (such as macs, gpu_cores or sources) are arrays; empty values are null.

{
  "name": "Which open model runs on which Mac, and how fast (estimates)",
  "id": "fit",
  "version": "2026-10-19",
  "dateModified": "2026-10-19",
  "url": "https://mlx.app/data#fit",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "attribution": "MLX Network, https://mlx.app/data",
  "rows": [
    {
      "mac": "m1-8gb",
      "chip": "Apple M1",
      "memory_gb": 8,
      "bandwidth_gbs": 68,
      "model": "qwen3-0.6b",
      "model_name": "Qwen3 0.6B",
      "version": "8-bit",
      "bits_per_weight": 8.5,
      "weights_gb": 0.6,
      "needs_gb": 2.1,
      "fit": "fits",
      "group": "Runs well",
      "below_4_bit": false,
      "tok_s_low": 51,
      "tok_s_high": 91,
      "tok_s_ceiling": 113
    }
  ]
}

#CSV

A CSV file has a header row with the column names, \r\n line ends and RFC 4180 quoting. List values are joined with ; , and empty values are empty cells. The content type says header=present.

mac,chip,memory_gb,bandwidth_gbs,model,model_name,version,bits_per_weight,weights_gb,needs_gb,fit,group,below_4_bit,tok_s_low,tok_s_high,tok_s_ceiling
m1-8gb,Apple M1,8,68,qwen3-0.6b,Qwen3 0.6B,8-bit,8.5,0.6,2.1,fits,Runs well,false,51,91,113
m1-8gb,Apple M1,8,68,qwen3.5-0.8b,Qwen3.5 0.8B,8-bit,8.5,0.9,2.4,fits,Runs well,false,34,60,76

#Columns

The manifest describes every column in full; in short:

macs: mac (id, as in mlx.app/mac/<mac>), chip, chip_id, generation, tier (base, pro, max, ultra), memory_gb, bandwidth_gbs (as Apple publishes it; empty where Apple publishes none), bandwidth_gbs_used (what the estimates use), gpu_memory_default_gb (what macOS lets the GPU use by default), gpu_cores, macs (the Macs the chip ships in), note, page, sources.

models: model (id, as in mlx.app/run/<model>), name, maker, hf_repo, params_b, active_params_b (parameters read per generated token), architecture (dense or moe), context_tokens, license, released, modality, precision, repo_weights_gb, native_gb and native_bits (for models released in a low-bit format), offload_gb (lookup tables a runtime can leave on disk), mlx_build, page, sources.

fit: mac, chip, memory_gb, bandwidth_gbs, model, model_name, version (the highest precision that fits), bits_per_weight, weights_gb, needs_gb (weights plus room for context and the runtime), fit (fits, tight, no), group (the label the pages show), below_4_bit, tok_s_low and tok_s_high (the estimated writing speed range), tok_s_ceiling (memory bandwidth divided by the bytes read per token).

#What's left out

The Artificial Analysis Intelligence Index shown on mlx.app/compare isn't in these files: its terms reserve redistribution. The llama.cpp measurements the hardware catalog cites are linked from the pages, not copied.

#Licence

The data is CC BY 4.0: use it, change it and share it, commercially too, as long as you credit MLX Network, https://mlx.app/data.

#Errors