ReflectionAI Unveils First Open-Weight Model,

U.S. AI startup ReflectionAI unveiled Beam, its first open-weight model. ReflectionAI announced the Beam launch on the 5th, local time, and said it plans to distribute the model weights under the Apache 2.0 license later this month. The company also plans to release a technical report, a model card, and developer tools.

Beam is a model aimed at coding, reasoning, and AI agent tasks. ReflectionAI said it is positioning Beam to compete with major Chinese open-weight models.

Beam is a text-based model that uses a sparse mixture-of-experts (MoE) architecture. It has 501 billion total parameters and 23 billion active parameters during token processing. The pretraining data used 23.8 trillion tokens, and the model can handle a context window of up to 1 million tokens.

NVIDIA GB300 GPUs were used to train the model. A total of 6,144 NVIDIA GB300 GPUs were used for pretraining, and 10,500 NVIDIA GB300 GPUs were used for reinforcement learning. More than 100 million rollouts were generated during reinforcement learning.

ReflectionAI highlighted Beam's competitiveness by comparing its performance and compute efficiency with major Chinese open-weight models. The company compared Beam's performance with those models and said some of its reasoning and coding benchmark results are similar to Z.ai's GLM-5.2. It also said Beam's results are close to Alibaba's Qwen3.8-Max.

ReflectionAI also emphasized compute efficiency. The company said Beam's performance on difficult reasoning benchmarks is similar to GLM-5.2, and that the amount of compute required for Beam's inference is about one-third to one-quarter that of competing models.

However, those performance figures are based on ReflectionAI's own evaluations. In addition, some models, including Moonshot AI's Kimi K3, showed stronger performance than Beam.

ReflectionAI plans to release Beam's weights later this month after final red-team testing. The company is targeting enterprise and institutional demand for AI deployments based on open-weight models, and those organizations can run the models on their own infrastructure and adjust them for their own use cases.

Shinsegae Group, which is pursuing sovereign AI infrastructure deployment in South Korea, signed an MOU with ReflectionAI in San Francisco in the U.S. last March. The MOU aims to form a strategic partnership to build a Korean sovereign AI factory. The two companies agreed to cooperate on building AI infrastructure using ReflectionAI AI models and NVIDIA GPUs.

Separately, discussions are under way with Shinsegae Group on building an AI data center (AIDC). Gangwon State and Chuncheon City are in talks over attracting a Shinsegae Group AI data center to the Suseol energy convergence cluster in Dongmyeon, Jinae-ri. However, the scale of the AI data center investment has not been finalized, and the construction schedule has also not been finalized.

ReflectionAI was founded in 2024. Its founders are CEO Misha Laskin and Ioannis Antonoglou, both former Google DeepMind executives. Investors in ReflectionAI include NVIDIA, Sequoia Capital, and Lightspeed Venture Partners.

The company secured computing resources based on NVIDIA GB300 GPUs via SpaceX and Nebius, and is training follow-up models with those resources. Based on this, the company plans to continue competing in the open-weight model market, starting with Beam, and the leading players in the open-weight market are Chinese companies.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407782

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407782