Reflection AI unveils Beam, a 501-billion-parameter open-weight model built to compete with China's GLM-5.2

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. On October 5 it introduced Beam, its first open-weight model. Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active per token, and a 1-million-token context window. It was pretrained on 23.8 trillion tokens of curated web data and proprietary licensed datasets. Access is limited to an early-access program for now. Reflection says it will publish the weights under Apache 2.0 later this month, together with a technical report, a model card, and tools for evaluation and fine-tuning.

Reflection disclosed its compute budget. It pretrained the base model on 6,144 Nvidia GB300 GPUs in under four weeks. The reinforcement-learning stage then ran on 10,500 GB300s for four weeks and produced more than 100 million rollouts. Wccftech reports that the RL environment ran 46.4 million sandboxes per day. The post-training budget was larger than the pretraining budget. That matches how labs now build agentic coding models, where most of the spending goes to RL against executable environments.

Reflection's own scorecard gives Beam 80.9 on SWE-bench Verified, 80.1 on Terminal-Bench 2.1 and 97.8 on the American Invitational Mathematics Examination (AIME) 2026. Against Z.ai's GLM-5.2, Beam scores 65.5 to GLM-5.2's 62.1 on SWE-bench Pro v1, and 80.1 to GLM-5.2's 81.0 on Terminal-Bench 2.1. AIME is near its ceiling at this tier and tells readers little. On the agentic coding evaluations that separate models, the published numbers show a tie with GLM-5.2. They do not show a lead.

The efficiency claim is the basis of the launch, and it is the weakest part of the case. Reflection says Beam matches GLM-5.2's reasoning scores with three to four times less inference compute. The Next Web notes that the figure is an estimate across DeepSWE, Humanity's Last Exam and Terminal-Bench 2.1, and that Reflection itself calls it an approximate comparison rather than a measured cost. Nobody has verified any of the scores independently. A 23-billion-active design does reduce per-token FLOPs. Serving still needs about 501 GB for the weights alone at 8-bit precision, so Beam remains a multi-GPU deployment and will not run on a single node.

The framing is geopolitical. Chinese-language technology media reported the launch as a US challenge to DeepSeek and Kimi, and Taiwanese outlets asked whether Beam can compete with GLM and Qwen. Reflection belongs to Nvidia's Nemotron Coalition of US open-model builders. It has signed a deal with South Korea's Shinsegae Group, a conglomerate, for a 250-megawatt data center that will run Korean-language models. It is reported to be raising $2.5 billion at a $25 billion pre-money valuation, meaning before the new money is counted. That compares with the $8 billion valuation at which Nvidia invested about $800 million.