FireFlash released: 64 B300s completed training in less than 5 days, reducing single-task costs by 30 times.

FireFlash completed training using 64 B300 GPUs in less than 5 days, matching the performance of Fable 5 and GPT-6Sol, with a 30-fold reduction in single-task cost. The training cost for long-horizon agents is high; a 3-hour sample training with hundreds of rounds of calls requires thousands of GPUs, and FireFlash significantly lowers this barrier.