Long-Horizon Agentic RL

Researchers released FireFlash, a 300B terminal and coding agent trained via full-parameter reinforcement learning from GLM-5.3-Flash, reaching 46.2% Pass@1 on Terminal-Bench 4.0, on par with Fable 5 and GPT-6 Sol while costing over 30x less per task than Fable 5 at maximum reasoning effort. The entire run used just 64 NVIDIA B300 GPUs for under five days, and the team published the full long-horizon agentic RL recipe.