Reflection AI open-sources Beam 1.0: a 501B MoE that triples agentic coding score after a 40-step RL run
rhythmrg · x · 2026-10-17
Reflection AI released Beam 1.0 under Apache 2.0 — a 501B-parameter MoE with 23B active, built with 40+ partners. The release includes weights (BF16/FP8/NVFP4), tokenizer, reference implementation, a DFlash2 drafter, and Mirror CLI, the open-source harness it was trained on; weights are on Hugging Face. Applied Compute ran a brief RL post-training on AC2 that tripled Beam's score on a challenging long-horizon agentic coding task in just 40 steps (0.07 → 0.215), and offers day-0 inference and post-training support.
Related event: Reflection AI Open-Sources Beam-1.0-501B-A23B with 40+ Ecosystem Partners(40 posts)→
More from coding & agent
- Backend migration from Python to Go now takes one engineer just one week — rakyll · 2026-10-17
- Audio founder says Claude one-shot decompiled rival algorithms, moat gone — ChrisGPT · 2026-10-17
- Google's playbook for agent evals: four layers of checks beyond the vibe check — rseroter · 2026-10-17
- AgentForesight: a 7B online auditor that catches multi-agent failures mid-run — jiqizhixin · 2026-10-17
- Qwen3.8-27B shows a weird write-delete loop in /tmp before every web research task — Training-Respect8066 · 2026-10-17
- Dev gives his agent an alarm tool to wake him during overnight /goal runs — zack_overflow · 2026-10-17