The company points users to a Kalshi cookbook and offers both Managed Training and a Serverless Training API.
Fireworks trained a LoRA adapter in one pass over 20,051 historical Kalshi questions from 6,461 events, using K-way cross-entropy over each question's answer choices. It reports 64.8% accuracy on its full evaluation set, versus 56.4% for base Glimmer and 45.9% for Jev. On 113 topic families excluded from training and calibration, tuned Glimmer reached 58.3%, compared with 53.2% and 47.1%.
The company separately calibrated the tuned model's confidence. At predictions of at least 80% probability, it reports 86.8% accuracy, while 14.4% of the full test set and 6.1% of unseen-topic questions met that threshold. Its historical backtest turned a hypothetical $1,000 into $1,071 after modeled fees. The source does not establish independently replicated results, real-world trading performance, or an equivalent comparison: base Glimmer received neither task-specific training nor confidence calibration.