Soup CLI
Prepare data, configure model training, evaluate checkpoints, and ship through one CLI.
Train specialized small models through an agent-driven workflow

Freesolo provides tools for adapting smaller AI models to specific tasks through post-training. Its Flash package is designed for coding agents such as Claude Code, Cursor, and Codex to operate: a developer describes the target behavior, prepares a training environment, reviews a quoted run, and works toward a specialized model.
The workspace organizes environments, training runs, and resulting models, with support for supervised fine-tuning and reinforcement-learning workflows. Trained weights can be downloaded in standard formats or served through a compatible endpoint. This approach is intended for teams exploring task-specific latency and cost tradeoffs; the usefulness of the resulting model still depends on the training data, task, and evaluation.
Tried it? Share a useful thought or ask a question.
Good conversations start with a hello.
Sign in to comment