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

Reflection announced Beam, a text-only mixture-of-experts model with 501B total parameters and 23B active parameters, aimed at coding, agentic and scientific work. The company said it was trained from scratch and that full weights under Apache 2.0 are due this month.

Reflection team posts cited 23.8T pretraining tokens, partly produced by an OCR pipeline over hundreds of millions of PDFs. They described a stable RL/OPD run on 10,000 GB300s with more than 100M rollouts across roughly 1M tasks. A summary of Reflection's claims listed 80.9 on SWE-bench Verified, three to four times the inference efficiency of GLM 5.2, and four weeks each of pretraining and RL on about 10,500 GB300s. A tech report and open-source integrations were promised.

Axios reported the launch ahead of time, saying Reflection pays $150M per month for Colossus compute plus a $1B Nebius deal, and that other unnamed US labs will ship open models this month. Artificial Analysis said it had early access and expects Beam to be among the most token-efficient open models for its intelligence.

Independent analyses were mixed. Elie Bakouch estimated only about 12% BF16 MFU in pretraining, read the architecture as 3:1 interleaved global/sliding-window attention, and noted better held-out code perplexity than DSv4. Teortaxes called Beam an iso-FLOP replication of DeepSeek V3 and inferred roughly 1.3B RL sandboxes over four weeks, with up to 170K running at once. Observers placed Beam around GLM-5.2 level and below DSv4 Flash on some benchmarks. Nathan Lambert grouped it with Nvidia and Thinking Machines as strong US releases that still trail Chinese counterparts.

The report also covered other releases: Aleph Alpha's Kolibri (78B total / 3.46B active, Apache 2.0, German and English), Reka's Rho-1 19B omni model trained on 320 H100s in about three months, Command Code's Agr (31B) and Agr-flash (360M) decision models, and Upstage's Solar Mini 4 (35B / 3B active, 512K context), free on Nous Portal for two weeks.