FireFlash: 64 B300 GPUs, 5 days of RL matches frontier coding agents at 30x lower cost
kaiwei_chang · x · 2026-11-03
The FireFlash team post-trained a 300B-scale long-horizon coding agent from GLM-5.3-Flash using just 64 NVIDIA B300 GPUs for 5 days, scoring 46.2% on Terminal-Bench 4.0 (+10.4 over baseline) while costing 30x less per task than Fable 5 Max.
- Gains transfer to DeepSWE v1.1 (68.8, +5.4) and SWE-Bench Pro Verified (63.3, +5.1)
- The model sits on the performance-cost Pareto frontier, near GPT-6 Sol (49.4%) at an order-of-magnitude lower price
- Three keys: curated agentic RL environments with reliable rewards, an RL recipe built for long/stale/low-precision rollouts, and a coding harness co-designed for training and inference
The full RL recipe is being released openly.
Related event: FireFlash builds frontier coding agent with just 64 B300 GPUs in 5 days(2 posts)→
More from coding & agent
- Watching three AI agents interview each other in Slack to write docs — aniketmaurya · 2026-11-03
- AI Products Still Feel Single-Player; Team Bots Point to Multiplayer Human-Agent Future — joonasvirtanen · 2026-11-03
- Stanford launches CS329Z on Engineering AI Agents, taught by SWE-agent creator — stanfordnlp · 2026-11-03
- Podcast: opencode's thdxr on open source, harness vs model, B2B vs B2C AI — yenkel · 2026-11-03
- 86.2 Agent Hours With Opus 5.5 Produced a Hyperpop Music Video — apples_jimmy · 2026-11-03
- Matt Shumer shares "Manager Loop": the orchestration trick behind super long-horizon agent builds — mattshumer_ · 2026-11-03