Aden Experimental

by Jeff

Aden Experimental is what we call our roasting models while they're in research. Each one is a small open-weight model trained to do one job: given a green coffee and a roaster, write a roast program in the Roast DSL that the roaster can run.

They aren't released. For now we're not publishing the weights or offering the models as a product.

How they're trained

• A small base. The current model is fine-tuned from Qwen3.5-9B, small enough to serve cheaply.
• Learning the language. Supervised fine-tuning on conversations drawn from the DSL's documentation, its machine registry and real roast programs, so the model knows every machine's channels, ranges and units.
• Learning to roast. Reinforcement learning against roast simulators fitted to recorded roasts. The model writes a program, the simulator roasts it, and the curve is graded much as Roast Bench grades it.

What we've learned so far

• The language is the easy part. In the first version, every program the model finished compiled and was legal on its machine. Syntax, channel names and units were learned.
• Finishing is harder. Early versions sometimes never stopped, writing hundreds of tiny steps until they ran out of room. That came from the model's thinking mode, not the format, and training against the simulator is how it's being fixed.
• Small can roast. On Roast Bench's Aillio items, the current model's single answer has a median of 97.6, on par with much larger general models and far ahead of its base model's 19.8. Its misses are mostly roasts that stall before first crack.

What's next

More roasters, as their simulators pass their fits. Then the half of roasting the benchmark can't grade yet: hitting the roast degree and development the roaster asked for.