README / README.md
ur-dad-matt's picture
Embed logo + quickstart + tier matrix
738e1ba verified
metadata
thumbnail: https://huggingface.co/Outlier-Ai/README/resolve/main/assets/thumbnail.png

Outlier

Outlier

Frontier AI on hardware you own.

Nano (1.7B) · Lite (7B) · Compact (14B) · Max (32B)


What is Outlier

A family of small, fast, local language models designed for the hardware people actually own — Apple Silicon Macs, consumer NVIDIA GPUs, and CPU-only laptops via GGUF.

Four tiers from 1.7B to 32B, shipped in three formats:

Format Use case Runtime
MLX 4-bit Apple Silicon Macs Outlier Desktop app, mlx_lm
AutoAWQ 4-bit NVIDIA GPUs (datacenter) vLLM, transformers+awq
GGUF Q4_K_M / Q5_K_M CPU-portable, any OS llama.cpp, Ollama, LM Studio

Quickstart — Apple Silicon

pip install mlx-lm
python -c "
from mlx_lm import load, generate
model, tok = load('Outlier-Ai/Outlier-Lite-7B-MLX-4bit')
print(generate(model, tok, prompt='Hello', max_tokens=100))
"

Quickstart — Ollama

ollama pull hf.co/Outlier-Ai/Outlier-Lite-7B-GGUF
ollama run hf.co/Outlier-Ai/Outlier-Lite-7B-GGUF "Hello"

Desktop app

Outlier ships a native Mac desktop app with streaming chat, markdown rendering, session history, model switching, and a full settings panel. See outlier.host (coming soon) or download from the Releases page.

Research

Three provisional patents filed April 2026 (61 claims total) on the Path F+ BTX merge architecture that produces the tier family from base models. Non-provisional consolidation due April 2027.

License

Apache 2.0 on all models. Base weights inherit Qwen 2.5's license. Post-training artifacts released under the same terms.

Contact

mattkerr09@gmail.com — Matt Kerr, solo founder, Michigan.