Models named for the job, not the parameter count.
Each model is an open-weight model chosen for one kind of work. Settings shows its download size and the memory it needs before you get it, and nothing downloads until you ask. Every model links to its benchmark results.
The Trixe lineup
Small and fast
Lighter models for older computers and instant replies. Nano is never picked automatically: in testing, its answers weren't reliable enough.
Power tiers
Without Trixe, LCOA runs the Qwen3 family through Ollama and suggests a tier for your hardware. There is also Reflex, which loads no model at all and answers only what plain code can.
Triangle Auto
Leave the choice to Triangle and it scores each prompt on quality, task fit and speed, then picks the best model that fits in memory. It never downloads anything by itself, the chat shows why it chose what it did, and it keeps the smallest models away from code, math and anything that touches your files. Try it on the home page.
Worms
Ten tiny specialists, trained from scratch
Each Worm has 19.3 million parameters and one job. They are experiments: in testing, several wrote summaries that missed the point, and the Math Worm got 17 + 25 wrong. So LCOA calls one only when you ask for it, and the main model checks what it says against the source before using any of it.
- Frontend
- Backend
- Bug Finder
- Math
- English
- File Reader
- Code Reader
- Test Writer
- Data Analyst
- Planner