METHODOLOGY

prices:
  what: official maker list prices, MSRP and SEP from the maker's own store or announcement (Framework, Apple, NVIDIA incl. NVIDIA Marketplace, AMD, Dell, Lenovo, Corsair, Crucial, Minisforum, Beelink, GMKtec); AMD Ryzen AI Halo uses Micro Center, AMD's only US seller
  not_used: Keepa, CamelCamelCamel, PCPartPicker (terms bar scraping or republishing); retailer pages until partner APIs are set up
  history: official announcements plus Internet Archive (Wayback) snapshots of the maker's own store page
  rule: no price is invented; every record stores source URL and fetch date
  public_vs_paid: public pages show the current price, the last 7 days of records and the current source; every older record and its source is in the paid YAML / CSV export
  listed_products: a product is listed once it has 2+ recorded prices, or 1 record and a launch in the last 180 days (marked *NEW*)
changes:
  windows: 1d, 7d, 14d, 30d, 90d, 180d, plus change since the first record (since first)
  baseline: the last recorded price on or before the lookback date, accepted only if recorded within min(14, max(3, N)) days before it
  since_first: latest price vs the first recorded price; shown with the month of that first record
  price_hikes: count of recorded increases between consecutive records; 'latest step up' = size of the most recent increase and the first date we saw it (a Wayback gap means it may have happened earlier)
  out_of_stock: shown only when the maker page itself said out of stock at our latest record
  trend_charts: shape of recorded prices since the first record, carried forward between records; no interpolation between different prices
  marks: — = no record close enough to that date · new = first record is after that date
  index: equal-weight average of the products that have a value for that window; n = how many
canrun:
  models: a fixed list of current open-weight models, from 12B to 120B parameters, plus Llama 2 7B as the llama.cpp reference benchmark
  model_sizes: GGUF file size from the Hugging Face API for each quant (Q4_K_M or Q4_0, Q8_0, BF16/F16, or the native MXFP4)
  kv_cache: fp16 KV cache at 8,192 tokens from each model's config.json: 2 × KV heads × head dim × 2 bytes × tokens per attention layer; sliding-window layers count only their window; Mamba / linear-attention layers keep a small fixed state and are not counted
  overhead: +10% of the file size for compute buffers and runtime
  usable_memory: discrete GPU: VRAM − 1 GB · unified memory (Mac, Strix Halo, GB10): 75% of total · DDR5 kits: 75% of capacity, CPU-only
  fits: YES = needs ≤ usable · TIGHT = needs ≤ total memory (may work with a raised GPU memory limit or shorter context) · no = larger than total
  largest: CAN RUN (largest) = the model with the most parameters that fits with YES
  speed: token-generation results published in llama.cpp benchmark threads; chip-level rows are labelled as such; no estimates
  bandwidth: maker spec pages; DDR5 kits computed as MT/s × 8 bytes × 2 channels
emphasis: rises are shown in red, bold, upper case and underlined; falls in green
affiliate: store links may carry affiliate tags; tags never change which product a guide ranks first