aquaduck / Llama-3.2-3B-Instruct-MLX

Model Card for aquaduck/Llama-3.2-3B-Instruct-MLX

Pinned 4bit MLX of Llama-3.2-3B-Instruct ( meta-llama/llama-3.2-3b-instruct ), plus midpoint layer shards for staged / multi-node loading (Aquaduck Arc mlx-package-v1). The shard files are not a new quantization. They are contiguous midpoint packages cut from the full 4bit MLX in this repo.

Model lineage

meta-llama/llama-3.2-3b-instruct └── quantized → meta-llama/Llama-3.2-3B-Instruct (4bit) └── full MLX + midpoint shards → aquaduck/Llama-3.2-3B-Instruct-MLX (this repo)

• Base weights: https://huggingface.co/meta-llama/llama-3.2-3b-instruct (other)
• Quantization source: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct (tag 4bit)
• This repo: full 4bit MLX ( model.safetensors ) and midpoint MLX shards

Model Details

Model Description

• Hosted by: Aquaduck (hosting and layer packaging only; base model by Meta; MLX quant by mlx-lm)
• Shared by: Aquaduck AI
• Model type: Causal language model (Llama-3.2-3B-Instruct), MLX 4bit
• Language(s): Multilingual (same as base)
• License: other (inherits from meta-llama/llama-3.2-3b-instruct)
• Finetuned from model: N/A — not a fine-tune
• Derived from: meta-llama/Llama-3.2-3B-Instruct ← meta-llama/llama-3.2-3b-instruct

Model Sources

• Base model card: https://huggingface.co/meta-llama/llama-3.2-3b-instruct
• Quantized MLX source: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct

Files

• Total layers: 28
• Valid split boundaries: 14 Filenames use exclusive end indices (layers-{start}-{endExclusive}/model.safetensors).

Uses

Direct Use

• Full model.safetensors : standard single-file 4bit MLX (mlx-lm-compatible). Use this for single-node / local runs.
• layers-*-*/model.safetensors : Aquaduck / Arc staged loading only. These are not drop-in complete models for stock mlx-lm. Use the base model’s chat template (including thinking / instruct modes as documented on the base model card); other formats will not work correctly.

Out-of-Scope Use

• Expecting any one shard to run as a complete model
• Treating this repo as a new training run or re-quant
• Uses prohibited by the other license or the base model’s model card guidance

Bias, Risks, and Limitations

Same capabilities, biases, and risks as meta-llama/llama-3.2-3b-instruct. 4bit quantization can degrade quality vs. the original higher-precision releases. Layer sharding does not change weights beyond packaging.

Recommendations

Follow the base model’s docs for chat template, thinking vs instruct modes, and sampling. Prefer model.safetensors in this repo when you do not need staged loading.

How to Get Started

These files are meant to be loaded automatically by the Aquaduck desktop app .

1. Download the Aquaduck desktop app and sign in.
2. Devices connected to the internet will receive a model assignment from the model catalog ( meta-llama/llama-3.2-3b-instruct ).
3. Download the model from the Home view. The app will:
   • download only the assigned file from this repo (full model.safetensors or one midpoint half)
   • keep that stage ready for serving You do not need to pick files by hand, but you may for local serving. Assignment and download are driven by model catalog metadata. The full model.safetensors is a standard 4bit MLX. The layers-*-*/model.safetensors files are not.

Training Details

No training. Weights come from Meta; 4bit MLX from meta-llama/Llama-3.2-3B-Instruct; this repo hosts that MLX and (when split) packages it into midpoint layer shards.

Evaluation

No separate evals for the hosted MLX or shards. See meta-llama/llama-3.2-3b-instruct.

Technical Specifications

• Architecture: Llama-3.2-3B-Instruct (~3.2B params, GQA (24 Q / 8 KV heads), 28 layers, hidden dim 3072)
• Quantization: 4bit
• Packaging: pinned full 4bit MLX; optional Arc midpoint shards ( layers-{start}-{endExclusive}/model.safetensors )
• Package format: mlx-package-v1
• Split: 2 stages at layer 14 (maxStages: 2)

Citation

@misc{llama323binstruct,
    title  = {Llama-3.2-3B-Instruct},
    author = {Meta},
    year   = {2026},
    url    = {https://huggingface.co/meta-llama/llama-3.2-3b-instruct}
}

Credit:

• The MLX quantization source ( https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct )
• mlx-lm ( https://github.com/ml-explore/mlx-lm ) for MLX support

Attribution

Quantized MLX ingested from meta-llama/Llama-3.2-3B-Instruct . Original weights: meta-llama/llama-3.2-3b-instruct . Redistributed under the base model's license.

Hosted by Aquaduck .

Model Card Contact

Aquaduck AI — https://huggingface.co/aquaduck