mlx-serve/Community · Tier list

The local LLM tier list.

Community-voted rankings of the LLMs that run locally on Apple Silicon — auto-updated from Hugging Face, filtered to your Mac's unified memory. One Google account, one vote per model.

My Mac has
Must have
Loading votes…
S
A
B
C
D

Not yet ranked

The first community vote puts a model on the board. Same hardware filters as above; rows hold their position while you vote.

ModelParams~SizeRuns on↓ / moTagsYour vote
How it works

Honest rankings, tiny rules

Voting

One account, one vote per model

Sign in with Google, then ▲ or ▼ any model. Click the same arrow again to take your vote back, or the other one to flip it. Your votes are keyed to your account — refreshing, reinstalling, or switching devices can't double-count you.

Tiers

Wilson score, not raw sums

Tiers come from the lower bound of the Wilson confidence interval on the up-ratio — a first vote enters a model cautiously (C), and it needs consistent upvotes to climb toward S. Ten early fans can't out-rank a hundred mixed reviews.

Hardware

The filter is about fitting

Each model lists its approximate quantized footprint; the RAM filter hides anything that wouldn't leave your Mac room to breathe (weights ≲ 75% of unified memory). Votes are global — the filter changes what you see, not the scores.

Privacy

No emails stored, ever

Votes are stored against an opaque account id in Firestore. The public data is just model → up/down counts; your name and email never leave the sign-in widget in your own browser.

Found your model?
Run it in one command.

mlx-serve run <model> downloads it straight from Hugging Face and drops you into a chat — or grab the Mac app.

Download for Mac View on GitHub