Reflection AI Beam: New 501B Open-Weight Model
Reflection AI announced Beam, a 501B open-weight model trained on 10,500 GPUs to challenge Chinese labs. Public weights arrive by late October 2026.
Former DeepMind researchers trained a 501B sparse architecture on 10,500 GB300 GPUs, backed by rented compute at xAI’s Colossus 2.
Reflection AI Beam entered the frontier AI race on Monday as a 501-billion-parameter open-weight model built from scratch for reasoning and agentic workflows. Backed by Nvidia and valued at $25 billion, the US startup is positioning the release as a direct Western challenge to the open-source lead held by Chinese laboratories. The company plans to release full model weights under a permissive Apache 2.0 license by late October 2026.
How the Reflection AI Beam architecture works
Beam relies on a Mixture-of-Experts (MoE) design. Instead of running its full parameter count on every generated word, the system routes inputs to specialized sub-networks, activating 23 billion parameters per token. That setup lowers inference costs without reducing the model’s total capacity.
According to technical details published on the Reflection AI Blog, pre-training ran for four weeks across 23.8 trillion tokens. The team followed that phase with four weeks of reinforcement learning (RL), executing over 100 million rollouts across 10,500 Nvidia GB300 accelerators. Reflection AI claims an architectural context limit of 1 million tokens, though the current beta API restricts inputs to 256,000 tokens.
Billions in rented compute at Colossus 2 and Nebius
Reflection AI owns no physical data centers. The company relies entirely on rented capacity from external providers. Reporting by Axios revealed that the startup signed a $150 million monthly agreement in July to secure compute at the Colossus 2 supercluster, operated by SpaceX and xAI. The multi-year deal has an estimated ceiling of $6.3 billion through 2029, as noted in earlier coverage of SpaceX compute agreements by Axios, with a break clause requiring 90 days of notice after three months.
The startup also locked in compute capacity exceeding $1 billion with neocloud provider Nebius. That massive infrastructure spend explains how a venture formed in March 2024 assembled enough compute to train a tier-one model within weeks.
Western pushback against Chinese open-weight models
Over the past two years, the frontier of open weights shifted toward China, driven by releases from DeepSeek, Alibaba’s Qwen, and Z.ai’s GLM. Reflection AI founders Misha Laskin and Ioannis Antonoglou are pitching Beam as a Western counterweight to that ecosystem, according to reporting from Semafor.
The commercial strategy targets public sector and enterprise clients that want to run localized weights on private infrastructure. As outlined on the Reflection AI institutional page, the startup signed a memorandum of understanding with the US Department of Energy to deliver models across 17 national laboratories under the Genesis Mission. In South Korea, it partnered with retail group Shinsegae to develop a 250-megawatt sovereign computing facility.
What is verified and what remains unreleased
The model’s technical development and core infrastructure contracts are verified by corporate announcements and technical press reporting, including The Compute 100. However, the model weights are not yet hosted for public download on platforms like Hugging Face. Real availability remains tied to the release promised for later this month.
Performance claims also warrant scrutiny. The claim that Beam operates at inference costs three to four times lower than competing open models comes exclusively from Reflection AI’s internal benchmarks. Until independent research groups evaluate the final checkpoint on standardized public tests, those performance numbers remain vendor claims.