Reflection's 501B Beam model touted as open, but weights missing and generalization test challenged
bigdata · x · 2026-10-06
Reflection unveiled Beam, a sparse MoE model with 501B total / 23B active parameters, trained on 23.8T tokens with 100M RL rollouts on 10,500 GB300 GPUs, targeting coding and reasoning. Developers immediately asked for the Hugging Face repo; the answer is "coming soon." On Hacker News, its generalization test — a 180×90 world-map grid where Beam scored 95.5% coverage vs Opus 5's 92.5% and Fable 5's 97.8% — was challenged: the puzzle dates back to at least August 2025 on LessWrong, not "a few days old," and it's unclear whether web search was disabled.
More from Models
- Jeremy Howard: Google shipped the best model ever made; OpenAI its worst, same day — hugobowne · 2026-11-20
- Polymarket puts 83% odds on Google having the best AI model by year-end — Polymarket · 2026-11-20
- One year later, model naming is still a mess: the gpt-5.1-codex-max anniversary — heyneighbor · 2026-11-20
- Developer slams Grok models as 'garbage', shares coding agent's repetitive trial-and-error log — feregri_no · 2026-11-20
- Drex Router cuts small-game build cost to $1.26 vs $13.44 on Claude Opus alone — testingcatalog · 2026-11-20
- OpenAI slammed for AI slop: 'Just hire photographers to RLHF imagegen for a day' — flowersslop · 2026-11-20