FULL STORY
PufferLib 5.0 Trains Agents in One Second
Joseph Suarez released PufferLib 5.0, training RL agents in one second and setting new records on Craftax and NetHack benchmarks.
2026-09-14 ~ 2026-09-14 · 2 episodes · 8 posts
Episode 1 · PufferLib 5.0 Released: Train RL Agents in One Second (2026-09-14, 4 posts)
PufferLib 5.0 trains reinforcement learning agents in about one second, with 5x faster solving than 4.0 and peak throughput over 60 million steps per second. It abandons Python for under 10,000 lines of CUDA C implementing five-layer deterministic parallel training, with demo agents playable at puffer.ai.
- PufferLib 5.0 released: trains RL agents in under a second — jsuarez · 2026-09-14
- PufferLib team showcases 5 new results with playable agents online — jsuarez · 2026-09-14
- PufferLib 5.0 ditches Python: bitwise-deterministic async training in under 10k lines of CUDA C — jsuarez · 2026-09-14
- PufferLib 5.0 is ~5x faster than 4.0, peak throughput over 60M SPS — jsuarez · 2026-09-14
Episode 2 · PufferLib 5.0 Sets New RL Records on Craftax and NetHack (2026-09-14, 4 posts)
PufferLib 5.0 delivered a batch of RL breakthroughs: doubled SOTA on Craftax, a multi-task drone policy, self-play mastery of Robocode, and a 5x SOTA agent on NetHack that beats the best scripted bots.
- PufferLib 5.0: Five Coolest RL Results, from Multitask Drone Policies to Selfplay Robocode Champions — jsuarez · 2026-09-14
- PufferLib 5.0 Doubles SOTA RL Performance on Craftax, Reaching 71% Max Return — jsuarez · 2026-09-14
- RL agent beats hand-coded AutoAscend on NetHack with PufferLib 5.0, 5x prior SOTA — jsuarez · 2026-09-14
- PufferLib 5.0 RL Results: 19.5k NetHack Score with Random Classes, Multitask Drone Policies — j_foerst · 2026-09-14