Reflection launches Beam: 501B-param agentic open model trained with AdamC optimizer
andersonbcdefg · x · 2026-11-09
Reflection AI announced Beam, a highly efficient agentic open model with 501B total parameters and 23B active, claiming frontier reasoning efficiency and advances on the Western open frontier for coding and agentic tasks. Trained end-to-end from scratch; full weights release this month.
AdamC author aarondefazio confirmed Beam was trained with his AdamC algorithm, which fixes the late-stage gradient norm explosion in long LLM training runs (arXiv:2506.02285). Thinking Machines' blog also mentioned using it, indicating the optimizer is seeing broad adoption.
More from Models
- Reddit users puzzled: why is everyone still on Qwen3 instead of the newer 3.5 models — opUserZero · 2026-11-09
- AMD CEO Lisa Su: Meta Muse hit 10M activations in 45 days as AI adoption accelerates — rohanpaul_ai · 2026-11-09
- Running Ripple 65B Preview NVFP4 on SGLang with 2x RTX PRO 6000: full recipe — HankYeomans · 2026-11-09
- DSpine boosts Qwen3.8-27B to 94-132 tok/s in llama.cpp, 1.75x over DFlash2 — pbaylies · 2026-11-09
- Claude keeps refusing bilingual ebook translations over copyright, even for personal use — Accomplished-Ad-1321 · 2026-11-09
- Models can be trained to talk to each other in inscrutable neuralese text — Sauers_ · 2026-11-09