Reflection AI unveils Beam: 501B-parameter open-weight model to take on China's coding models

#What happened

Reflection AI, the US startup founded in 2024 by ex-Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, announced its first open-weight model, "Beam," on October 5. It is a text-only Mixture-of-Experts model with 501 billion total parameters, 23 billion active per token, and a 1-million-token context window — purpose-built for coding, reasoning, and agentic workloads.

#The numbers

Beam targets China's open-weight leaders: Z.ai's GLM-5.2 (744B total / 40B active) and Alibaba's Qwen3.8-Max. Reflection says Beam matches GLM-5.2 on advanced reasoning benchmarks while using one-third to one-quarter of the inference compute. Reported scores: 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench v2.1, 97.8 on AIME 2026, and 90.5 on GPQA Diamond. Training ran on 10,500 Nvidia GB300 GPUs for four weeks across 23.8 trillion tokens. One caveat: independent verification of these claims is still pending.

#Why it matters

Beam is aimed at enterprise and government users that need a non-Chinese open-weight option for coding agents — exactly the gap the open-weight agent ecosystem has been missing. The company is Nvidia-backed ($800M invested, about $4.7B raised at a roughly $25B valuation) and has locked GB300 capacity at SpaceX's Colossus 2 datacenter through 2029.

#Availability

The model is in final red-team evaluation. Weights and the technical report ship later this month under an Apache 2.0 license.

Sources:

• Reuters (Oct 5, 2026): https://www.reuters.com/technology/nvidia-backed-reflection-unveils-first-ai-model-take-chinese-open-models-2026-10-05/
• MarketWatch: https://www.marketwatch.com/story/this-new-ai-model-could-help-america-close-a-technological-gap-with-china-c1d74492
• BigGo Finance: https://finance.biggo.com/news/20b8c3d6-6450-409b-9e1b-47d2ca89c28f

Verified Oct 7, 2026