How we score

Does it fit?

On an Apple silicon Mac we count about 75% of the unified memory as room for the model. The system needs the rest. That rule comes from canirun.ai. On a graphics card we count about 90% of VRAM. On a Ryzen AI Max mini PC (the Strix Halo boxes) we count about 70%, because the operating system and drivers keep a larger share.

A model that needs more than that room is a red "Won't fit". We plan around a practical 4-bit file, which is how most people actually run these models. The file size is on the i icon.

Will it feel good?

For chat, speed is roughly the memory bandwidth divided by the gigabytes the model reads for each word. A mixture-of-experts model only reads its active slice, so it can feel much faster than its file size suggests. The estimate can be optimistic once a conversation gets long.

• Runs well: it fits with a few gigabytes to spare, and chat stays at about 12 tokens a second or better.
• Tight fit: it fits, and a long chat will feel cramped.
• Runs slow: it fits, and you will wait on replies or on each picture.
• Pictures turn amber after about 40 seconds. Video turns amber after about three minutes, and we still say local video is only short clips.

Bandwidths we use include Apple's published figures (M4 at 120 GB/s, M4 Pro at 273, M4 Max at 546, M5 Max at 460 or 614, M5 Ultra at 1,200, M6 at 153 or 170) and NVIDIA's figures (RTX 4090 at 1,008 GB/s, RTX 5090 at 1,792 GB/s).

Close to ChatGPT (July 2026)

Chat scores use a fixed yardstick: the July 2026 versions of ChatGPT, Claude, and DeepSeek. The sentence stays put when this month's leaderboard moves. Within 15 points is "close to". Further away is "a step above" or "a step below". Far under all three is "basic". We never show a percent of ChatGPT. Open the board.

Those boards test the full model. A local run almost always uses a compressed copy, a quantized file, usually 4-bit. The i icon says so. The compressed copy is a little less sharp than the score. Pictures, voices, transcripts, scans, and video are not on that chat board, so their line names a model card instead of a July 2026 chatbot.

Every number on the site names its source and date. If that date is more than 30 days before this build, the line says "may be outdated". The July 2026 yardsticks are kept on purpose, so they carry that label until we choose a new month.

Prices

Every price names its source and the day it was checked. Some are a store listing. Some are estimates, and those say so. The one Check price button is the live number. Macs and graphics cards go to Amazon with a placeholder tag (`localaibuddy-20` until you set your own). Mini PCs go to the GMKtec or Minisforum store, also with a placeholder code. Apple does not pay commission on Mac hardware, so the buy button does not send people to Apple.

Which pages search engines see

There is a page for each job, and a page for each job on each computer. Search engines are asked to index about 40 of the task-and-computer pages, the ones whose answer is actually different. The rest stay on the site, with a note that tells search engines to look at the closest distinct page.