ztxt2026
A context engine that makes small local models act bigger by feeding them less, better context.
- State
- paused
- Started
- 2026
- Links
- —
Local models on a laptop fall behind frontier models on coding work. The usual explanation is size. ztxt tests a different one: that most of the gap is context.
It compresses a project into flat key-value context under an 8,192-token budget, keeps lessons distilled from its own past decisions, and runs Qwen 3.5 27B locally on MLX. Every interaction is saved as a trace, so the traces can later fine-tune the model on its own way of working.
It's set up as an experiment, with five hypotheses and a timestamped decision log, so the result is worth something either way.
Log
Added a skill resolver and a coding-aware complexity classifier.
Connected the MLX bridge so Qwen 3.5 27B runs locally inside ztxt.
Wrote the experiment framework: five hypotheses about compressed context, self-improving lessons, and fine-tuning on traces.