FireFlash: An Open Recipe for Long-Horizon Agentic RL
Researchers introduced FireFlash, a 300B-scale model and open reinforcement learning recipe designed for long-horizon agentic tasks that achieves 46.2% on Terminal-Bench 4.0 while reducing costs by over 30x compared to proprietary models. This breakthrough demonstrates that high-performance agentic reasoning can be achieved on a modest hundred-GPU-scale cluster rather than requiring thousands of GPUs, significantly lowering the barrier for open-source AI development. The model reached the performance-cost Pareto frontier in under 5 days using 64 NVIDIA B300 GPUs by utilizing curated agentic RL environments and an optimized coding harness.
## BACKGROUND
Long-horizon agentic reinforcement learning involves training AI models to execute multi-step, complex workflows over extended periods rather than generating single-turn answers. Terminal-Bench 4.0 is a benchmark designed to evaluate autonomous agents on advanced command-line and software engineering tasks.