mlx-community on Hugging Face — Where to Pull, How to Trust

The trusted upload zone

huggingface.co/mlx-community is the curated organization on Hugging Face where the MLX community uploads converted, quantized models. As of 2026-05, it hosts thousands of MLX-format models — most popular instruct models in multiple quantization levels, vision-language models, audio models, and a long tail of fine-tunes.

The reason this matters is provenance. Anyone can upload a model to Hugging Face. The mlx-community org is run with curation — uploaders are vetted, conversion scripts are visible, and a core group keeps an eye on quality. When you pull a model from mlx-community/..., you're pulling from a trusted reference set, not a random upload.

How to evaluate any MLX upload (in or out of mlx-community)

Before you load a model from a random uploader, check four things on the Hugging Face model page:

Download count. A high count over the last 30 days is a real signal — many people have loaded this and presumably it works. New uploads with no downloads need more scrutiny.
The conversion script or command. Most reputable uploaders document exactly which command they ran (e.g. mlx_lm.convert --hf-path X --quantize --q-bits 4 --q-group-size 64). Reproducibility you can verify.
The source model link. A good upload links to the exact upstream repo it was converted from. "Llama-3.2-1B-Instruct" is too vague; meta-llama/Llama-3.2-1B-Instruct as a link is what you want.
The quantization config. Open the model's config.json on Hugging Face directly (the file viewer makes this trivial). Confirm quantization.bits, quantization.group_size, and quantization.mode match what the model card claims.

The two-uploads-one-name dilemma

Sometimes you'll see two MLX-community uploads for the same source model with the same quantization (e.g. two different mlx-community/Llama-3.2-1B-Instruct-4bit-style repos). The single rule for picking — prefer the upload from the most active maintainer with the more recent date. Browse the uploader's profile; recent activity across many models is the signal that they're maintaining their conversions as mlx-lm evolves. A 2024 upload by an inactive uploader will silently miss any post-2024 dispatcher fixes.

When to upload your own

If you've converted a model that isn't in mlx-community yet (especially with a mixed-precision recipe nobody has uploaded), push it back. Use --upload-repo in mlx_lm.convert (covered in lesson 2). The community runs on this — your upload helps the next person not redo the conversion.

Code

Quick metadata sanity check on a HF MLX model·python

# Read the model card metadata + config.json without downloading the weights.
from huggingface_hub import HfApi, hf_hub_download
import json

api = HfApi()
repo_id = "mlx-community/Llama-3.2-1B-Instruct-4bit"

# Model card metadata (download counts, tags, etc.)
info = api.model_info(repo_id)
print(f"Repo: {repo_id}")
print(f"Downloads (last 30d): {info.downloads}")
print(f"Likes               : {info.likes}")
print(f"Tags                : {info.tags[:8]}")
print()

# Config.json — the source of truth on architecture + quantization
cfg_path = hf_hub_download(repo_id=repo_id, filename="config.json")
with open(cfg_path) as f:
    cfg = json.load(f)
print("model_type     :", cfg.get("model_type"))
print("quantization   :", cfg.get("quantization"))

List MLX-format uploads from a specific uploader·python

from huggingface_hub import HfApi
api = HfApi()

# Browse all models from the mlx-community org
models = api.list_models(author="mlx-community", limit=20, sort="downloads", direction=-1)
for m in models:
    print(f"  {m.modelId:60} downloads={m.downloads or 0:>8}  likes={m.likes or 0}")

Exercise

Browse the mlx-community Hugging Face org for one model in each of three categories: an instruct LLM you'd use for chat, a quantized model in a precision you don't usually use (e.g. Q8 if you usually use Q4, or vice versa), and one mixed-precision upload. For each, do the four-point evaluation from the lesson body (download count, conversion command, source link, quantization config). Report which feels most trustworthy and why. Two sentences total.