Models

Helios

Local agent models for Apple Silicon. Not a hosted API.

If your agent runs in someone else's datacenter, your runtime ends at the edge of their availability, their policy, and their pricing. That is a strange foundation for software you are supposed to trust to do work on your behalf.

Helios is the local model line: post-trained MLX releases adapted from open base models for on-device agent work — practical engineering behavior, tool-use judgment, uncertainty calibration. Three sizes, all open weights, all served with mlx_lm on Apple Silicon.

We did not train these from scratch. They are open base models, adapted and packaged so a local agent stack — Union Street, or anything else — has something useful to point at when it does not need to phone home. ORSI is the runtime piece that lets an agent reach them.

python3 -m mlx_lm generate --model UnionStreet/Helios-Pika-1.0 --prompt "Who are you?"

python3 -m mlx_lm generate --model UnionStreet/Helios-Lynx-1.0 --prompt "Summarize this repo."