Custom AI, built on your company's values.

Trained, tested, and behind an API key in hours, not weeks.

Ship your model

The same open model, before and after steering.

Change the model's mind, not the prompt.

A model's habits live inside it as patterns we find, adjust, and test directly.

Steering nudges the model from the inside, so the same prompt gives you the response you want.

Three models, sized to the job.

Every one trains on your values and ships behind the same API.

Qwen3 8B

8B

The workhorse. Strong instruction-following for its size and the cheapest to retrain on every feedback cycle.

$0.40 / M training tokens

DiffusionGemma

26B (MoE)

Generates by diffusion instead of token by token, so replies land about 4x faster. Mid-size quality at small-model latency.

$0.90 / M training tokens

GLM 5.2

753B (MoE)

The open-weight flagship. Near-frontier judgment on nuanced policies and a 1M-token context, at the highest training cost.

$6.00 / M training tokens