Helios-Lynx-1.0

Helios Lynx 1.0

Helios Lynx 1.0 is a merged MLX checkpoint from Union Street AI, adapted from Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed.

This model is part of the Helios line of local-first agent models. The goal is a practical Apple Silicon model with stronger identity, tool-use judgment, uncertainty calibration, and constitutional behavior for long-running local agents.

It was not trained from scratch by Union Street AI. It is a post-trained adaptation of open model work, packaged here as a merged MLX release for local inference.

Intended Use

• local AI agents
• codebase navigation
• infrastructure reasoning
• long-context synthesis
• tool-use planning and recovery
• Apple Silicon inference experiments

Notes

This is an experimental release. It is intended for research and local agent work, not high-stakes deployment without independent evaluation.

The release includes the merged MLX weights and MTPLX runtime sidecar files from the base checkpoint so it can be used in MTPLX-aware local runtimes.

Example

python -m mlx_lm generate \
  --model UnionStreet/Helios-Lynx-1.0 \
  --prompt "Who are you?"