Reflection unveils Beam: 501B-param open model with 23B active, built for agentic efficiency
alexpolozov · x · 2026-10-06
Reflection AI has announced Beam, an open model with 501B total parameters and only 23B active, positioned as a highly token-efficient agentic model trained end-to-end from scratch, with full weights to be released this month.
Key points:
- The company claims Beam advances the Western open frontier on coding and agentic tasks
- Focus is on frontier-level reasoning efficiency with a small active parameter count
- Artificial Analysis has been given access for independent benchmarking; early indicators suggest Beam could be among the most token-efficient open models at its intelligence level
Related event: Reflection AI Open-Sources 501B MoE Model Beam(24 posts)→
More from Models
- Hugging Face researcher: encrypted reasoning is the worst thing to happen to LLM evals — _lewtun · 2026-10-06
- Redditor Post-Trains Yandex AliceAI-80B-A3B From Scratch on 3 Local V100s, Update #5 — jjusko20 · 2026-10-06
- Qwen3.8 Locally Loops on Design Talk Instead of Writing Code, 16GB VRAM User Seeks Fix — mintybadgerme · 2026-10-06
- EPFL professor recalls: ChatGPT scored 10/20 on last year's ML final exam — maksym_andr · 2026-10-06
- Rumored Anthropic November IPO raises questions about Fable 5.5 October timing — haider1 · 2026-10-06
- Dev measures app security in 'minutes until GLM escapes the sandbox' — current record: 2 — lucasmeijer · 2026-10-06