Research
FireFlash open recipe claims frontier-level agentic coding at much lower cost
A 300B-scale model called FireFlash reportedly scores 46.2% on Terminal-Bench 4.0, on par with Fable 5 and GPT-6 Sol at over 30x lower cost per task, trained in under 5 days on 64 NVIDIA B300 GPUs. The authors say the full recipe will be released, with open weights listed as soon.