[AINews] Reflection Beam - 501B-A23B American Open Model

Reflection spent more than a year talking up its coding ambitions and staying quiet. Now it has a real product: Beam, a text-only mixture-of-experts model with 501B total parameters and 23B active. It’s aimed at coding, agent work, and scientific tasks, and the full weights are due under Apache 2.0 this month.

The company says Beam was trained from scratch on 23.8T pretraining tokens, including text scraped through an OCR pipeline that scanned hundreds of millions of PDFs. It also describes a stable RL/OPD run on 10K GB300s, with more than 100M rollouts across about 1M tasks. That is not a casual demo. Reflection is trying to show it can do the grim, expensive part of frontier model building, not just talk about it.

The headline claim is an 80.9 score on SWE-bench Verified, along with 3–4x the inference efficiency of GLM 5.2. Reflection also says pretraining and RL each took four weeks on about 10,500 GB300s. Independent early reads are more restrained: Artificial Analysis expects Beam to be among the most token-efficient open models for its intelligence, while others place it around GLM-5.2 level and below DeepSeek V4 Flash on some benchmarks.

That gap matters. Beam is arriving into a crowded open-model month, and several observers still think the leading Chinese releases are ahead overall. But for the US side of the market, Reflection’s launch does something simpler and more important: it turns a stealthy lab into a shipping one. In this corner of AI, that alone is a statement.