Nvidia-Backed Reflection AI Launches Beam Model to Rival Chinese Open-Weight Systems
Reflection AI, a startup backed by Nvidia, has unveiled its first open-weight model, Beam, as it seeks to challenge the dominance of lower-cost Chinese AI models in coding and agentic tasks. The model is designed to compete directly with offerings such as Z.ai's GLM-5.2 and Qwen3.8-Max, which have gained traction for their efficiency and customization.
Technical Architecture and Performance
Beam utilizes a mixture-of-experts architecture, featuring 501 billion total parameters. To optimize performance and reduce costs, the model activates only 23 billion parameters for each task. This approach aims to provide a faster, more economical alternative to larger, denser models. In comparison, Z.ai's GLM-5.2 utilizes approximately 744 billion total parameters with 40 billion active parameters.
Reflection claims that Beam is competitive with leading Chinese models on advanced reasoning benchmarks while using significantly less inference compute. The startup, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, is positioning Beam as a workhorse for enterprises and sovereign nations looking to build localized AI factories.
Strategic Infrastructure and Market Positioning
Reflection has secured substantial computing resources to support its development, including a deal signed earlier this year with SpaceX for capacity at the Colossus 2 data center. The startup has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, reaching a $25 billion pre-money valuation.
By releasing Beam as an open-weight model, Reflection is entering a crowded field that includes Meta, Mistral, and Cohere, as well as U.S.-based open-model developers like Thinking Machines Lab. The company plans to release the model's weights and full technical documentation later this month, with distribution planned through major hyperscalers and neocloud providers.
Reflection AI was selected as a certified consortium member of the US Department of Energy's Genesis Mission across 17 National Laboratories and partnered with Shinsegae to build a 250MW sovereign AI cloud in South Korea. The company plans to release Beam's weights and evaluation stack under an open Apache 2.0 license later in October 2026.
What open weights would allow buyers to verify
For companies considering Beam, the decisive event is the actual publication of weights, licence terms and evaluation materials, not the launch presentation alone. Open-weight access can enable independent testing, private deployment and adaptation, but operators still need infrastructure, model monitoring and security review. A large total parameter count does not necessarily imply an equally large cost per generated token in a mixture-of-experts model. Claims about performance against Chinese competitors should be compared under equivalent hardware, benchmark settings and inference budgets before being treated as evidence of a lower cost of ownership.