Reflection AI Open-Sources Beam: 501B-Param Sparse MoE With 23B Active, ~$20–40M RL Training
Madisonkanna · x · 2026-10-06
Reflection AI introduced Beam, a highly efficient agentic open model with 501B total parameters and 23B active per token, trained end-to-end from scratch; full weights release this month.
- Efficiency comes from an RL length penalty discouraging unnecessary tokens plus a Sparse MoE architecture
- Trained with 100+ million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks—an estimated $20–40M
- Claimed to advance the Western open frontier on coding and agentic tasks
Related event: Reflection AI Open-Sources 501B-Parameter Model Beam(30 posts)→
More from Models
- Insider speculation: GPT-6.5 expected within 4 weeks, above Fable 5.5 level — VraserX · 2026-10-06
- OpenRouter's top 10 providers by monthly token consumption reveal real usage, not benchmarks — FinanceYF5 · 2026-10-06
- "There's no reason for models to struggle like this": debate over AI naming chaos — a1zhang · 2026-10-06
- Dev prefers GPT 5.6 Sol's chatty style over the silent GPT 6.x models — emjotde · 2026-10-06
- Does OpenAI's Watermarking Degrade Model Quality? Reddit Debates the KLD Cost — ResearchCrafty1804 · 2026-10-06
- Z.AI's GLM-5.3, a 744B-parameter MoE with 1M context, lands on Amazon Bedrock — Zai_org · 2026-10-06