Use reinforcement fine-tuning against a custom grader for verifiable tasks instead of labeling outputs

Fireworks AI·medium signal

Reinforcement fine-tuning is now GA on small reasoning models (o4-mini) and trains the model against a programmatic grader rather than labeled completions, which fits verifiable-reward domains like code, math, and structured extraction; the April 2026 updates added cheaper global training and GPT-4.1/-mini/-nano model graders. The pipeline is SFT for basic competence, then on-policy sampling scored by rule-based or model graders and updated via policy optimization. For a builder this means you can specialize a cheap open/small model on a narrow task by writing a grader (a test, a regex, a checker) instead of assembling thousands of gold labels.