Reflection AI, an American startup backed by Nvidia, unveiled Beam on October 5, 2026, positioning it as an open-weight model intended to compete with leading Chinese open models at a lower computational cost. Beam is a sparse Mixture-of-Experts (MoE) model with 501 billion total parameters, but only 23 billion are active per token, which Reflection AI states makes it more efficient for inference. The company claims Beam achieves benchmark results comparable to Z.ai's GLM-5.2 and approaches Alibaba's Qwen 3.8-Max on coding and agentic tasks, while requiring three to four times less inference compute than GLM-5.2. These benchmark comparisons, along with the efficiency claims, are currently based on Reflection AI's internal evaluations and have not yet been independently verified.
The launch of Beam is part of Reflection AI's broader strategy to build "AI factories" for enterprises and sovereign nations. This initiative aims to allow institutions to develop their own customized, local AI systems by training Reflection's models on their proprietary data. This approach offers organizations greater control over their intelligence layer, helps protect against vendor lock-in, and provides a more predictable cost structure for scaling AI workloads. Reflection AI emphasizes that this model allows data to remain in-house, addressing concerns for entities like hedge funds and trading firms that handle sensitive information.
Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, has focused on automating software development and building open intelligence. The company has secured significant compute commitments, including agreements with SpaceX and Nebius for access to Nvidia AI servers. Earlier this year, Reflection AI signed a memorandum of understanding with South Korea's Shinsegae Group to establish a 250-megawatt AI factory in the country, supported by the U.S. and Korean governments. This partnership exemplifies the "AI factory" concept in practice, combining Reflection models with customer data and secured Nvidia-backed compute.
Beam was pretrained on 23.8 trillion tokens from web, public, and proprietary licensed datasets, with the training completed in under four weeks using 6,144 NVIDIA GB300 NVL72 GPUs. The company also conducted a substantial reinforcement learning campaign over four weeks on 10,500 GB300 GPUs, generating over 100 million task attempts. Reflection AI reported that during this training, Beam's browsing performance improved even without explicit browsing tasks in the reinforcement learning mix, suggesting transfer across agentic domains. The model also features a controllable length penalty, which encourages efficient problem-solving with fewer tokens.
While Beam is currently undergoing final red-teaming and evaluations, an early version is accessible to a select group of users via a waitlist. Reflection AI has committed to releasing Beam's weights under an Apache 2.0 license later in October 2026. This release will include a technical report, model card, documentation, and the full stack necessary for running, evaluating, and fine-tuning the model. The company also plans to announce distribution partners and integrations with open-source libraries and harnesses at the time of the full release.