Reflection, a U.S. artificial intelligence startup backed by Nvidia, has introduced Beam, its first open-weight model built for coding, reasoning, and agentic workloads. The company plans to release the model weights under an Apache 2.0 license later this month, along with a technical report and developer tools. The launch positions Reflection as a new Western competitor in a market currently shaped by closed systems from Anthropic, OpenAI, and Google, and by strong open models from China.
Beam's Technical Foundation
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters. Reflection pretrained the system on 23.8 trillion curated tokens from the web, public sources, and proprietary licensed datasets. The company focused the training pipeline on source code, technical explanations, mathematics, and scientific knowledge to support downstream agentic coding tasks.
Competitive Performance and Efficiency
Reflection says Beam advances the Western open-weight frontier and is competitive with larger open models such as GLM 5.2, while approaching Qwen 3.8-Max on coding and agentic benchmarks. Although frontier open models like Kimi K3 still lead on raw capability, Beam's main advantage is lower inference compute. On advanced reasoning benchmarks, the model achieves scores comparable to GLM-5.2 while using three to four times less compute.
High-Compute Reinforcement Learning
Reflection made high-compute reinforcement learning a central scaling axis for Beam, using 10,500 NVIDIA GB300 GPUs over four weeks. The run generated more than 100 million rollouts with a maximum context length of 256,000 tokens and used approximately 1.3 billion sandboxes. The company sourced nearly one million high-quality coding, agentic, and STEM environments, and reported that capabilities continued improving with additional reinforcement learning compute.
Open Weights and the AI Factory Vision
Beyond the model release, Reflection is pursuing an 'AI factory' vision that gives institutions more control over data, software, and computing infrastructure. The startup is already testing the concept through a partnership with South Korea's Shinsegae Group to build a sovereign AI factory. The approach targets organizations with sensitive data, including hedge funds and trading firms, that want customized AI systems without relying fully on external closed services.
Compute Capacity and Market Pressures
To support training and inference, Reflection has secured substantial computing agreements with Nebius and SpaceX. The SpaceX arrangement is worth up to US$6.3 billion, with monthly payments of US$150 million from July 2026 through 2029 for NVIDIA GB300 capacity at the Colossus 2 data center near Memphis, Tennessee. The company is expanding capacity as Western organizations seek alternatives to strong Chinese open models for sensitive enterprise deployments.
What Comes Next
Beam is undergoing final red-teaming and safety evaluations, and Reflection is offering early access through a sign-up list. Later this month, the company will release the weights, technical report, model card, and developer artifacts under an Apache 2.0 license. Chief Executive Officer Misha Laskin has said that reaching the highest levels of model performance takes time, comparing the process to building rockets.
The introduction of Beam shows that Western open-weight development is accelerating, with efficiency and customizability becoming central competitive advantages. By pairing an open model with Nvidia-powered infrastructure and enterprise-focused AI factory deployments, Reflection is targeting institutions that want strong performance at lower cost and with greater control over sensitive data. Its success will depend on final evaluations and real-world adoption, but the launch adds a significant new option to the global open AI market.