AI Briefing
KOSign in

FireFlash: 300B Agentic RL Model Matches Fable 5 on Terminal-Bench at 30x Lower Cost

·2026.11.06 01:17

Key point

The 300B-scale model achieves 46.2% on Terminal-Bench 4.0 using only 64 NVIDIA B300 GPUs over five days.

Details

  • FireFlash, a 300B-scale model for long-horizon agentic tasks, scores 46.2% on Terminal-Bench 4.0, matching Fable 5 and GPT-6 Sol while costing over 30x less per task than Fable 5 at maximum effort.
  • Performance gains extend to software engineering benchmarks, including DeepSWE v1.1 and SWE-Bench Pro Verified.
  • The model reaches the performance-cost Pareto frontier in under 5 days using a cluster of 64 NVIDIA B300 GPUs.
  • Key technical factors include curated agentic RL environments for reliable signal, an RL recipe optimized for long, stale, and low-precision rollouts, and a tailored coding harness for training and evaluation.
  • The full recipe is being released, with open weights announced as coming soon.
reinforcement-learningterminal-benchagentic-rlmodel-efficiencynvidia-b300fireflash
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Microsoft ResearchMicrosoft Research Asia Releases Agent Lightning v1.0: A 3,500-Line Framework for Harnessed Agentic RLHacker NewsResearchers Defeat World’s Best Stratego Player With Low-Cost AI Model AtaraxosHacker NewsLongCat-2.5 Launches with Native Multimodality and 1.6T Parameters for Agentic CodingTLDR AIFigma's unreleased editing model, post-trained on Qwen 3.8 27B, beats top third-party model on internal evalsRedditAxion 1.0: Open Agent Foundation Model Trained with Simple Three-Stage RecipeRedditBeipo AI releases NorthCUA, an open-source computer-use agent family achieving state-of-the-art results