Nvidia-backed AI startup Reflection AI has launched its first open-weight artificial intelligence model, Beam, as the company seeks to compete with more capable and lower-cost Chinese models in software development and agentic AI.
Reflection said Monday that Beam is designed to perform coding and agentic tasks at a level comparable to leading Chinese open-weight models, including Z.ai’s GLM-5.2, while closing the gap with Alibaba’s Qwen3.8-Max.
The launch adds another U.S. contender to a rapidly intensifying competition over open-weight AI, where Chinese developers have gained ground by releasing models that can be downloaded, customized, and deployed at comparatively low cost.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Reflection said Beam contains 501 billion parameters, the variables that determine how an AI model processes information. However, only 23 billion parameters are activated for an individual task.
The architecture allows the model to use only a portion of its total network for each request, reducing the amount of computing required to generate an answer and potentially lowering both operating costs and response times. The approach is relevant for coding and AI agents, where models may need to execute long sequences of tasks rather than simply answer individual questions.
By activating a smaller portion of the model for each request, Reflection is attempting to combine the capabilities associated with a much larger model with the economics of a smaller one.
Beam enters a market being shaped by Chinese AI companies that have challenged the assumption that the most capable AI systems must also be among the most expensive to develop and operate.
Reflection said Beam is competitive with Z.ai’s GLM-5.2 on coding and agentic tasks and is approaching the performance of Qwen3.8-Max. GLM-5.2 has around 744 billion total parameters and activates approximately 40 billion for each task, according to the company. Beam therefore has a substantially smaller active parameter count even though its total parameter count remains large.
Inference costs are becoming a serious competitive variable in AI. As companies move from experimenting with chatbots to deploying AI agents that perform work continuously, the cost of running models can become as important as their benchmark performance.
An agent tasked with writing and testing software, navigating applications, or completing a multistep workflow may generate substantially more model calls than a conventional chatbot interaction. Lowering the computational cost of each task can therefore have a significant impact when such systems are deployed at scale.
The shift is also creating pressure on U.S. AI developers from China’s open-weight ecosystem. Chinese models such as DeepSeek and Kimi have gained attention for offering capabilities at lower costs while allowing developers greater flexibility to modify and deploy the underlying models.
U.S. technology companies are consequently competing on several fronts at once: raw model performance, inference economics, availability of computing capacity, and the ability to build ecosystems around their models.
For Reflection, the open-weight strategy places Beam directly in that competition rather than restricting the model to a closed API.
Reflection was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. The startup develops AI tools designed to automate software development, putting coding agents at the center of its business model.
Coding has emerged as one of the most commercially promising applications for advanced AI because software development can be decomposed into tasks that models can increasingly perform with limited human intervention. AI systems can generate code, identify errors, modify existing software and, increasingly, work through longer development workflows. That has made coding a natural testing ground for agentic AI, but also one where cost and reliability are critical. A model that can produce strong code but requires excessive computing resources may be difficult to deploy economically across large developer populations.
Beam’s architecture appears designed to address that problem by reducing the active computation required for each task while maintaining a much larger overall parameter base.
Reflection’s access to computing capacity is also becoming an important part of its ability to compete. Earlier this year, the company signed a deal with SpaceX for additional computing capacity at Colossus 2, the data center operated by Elon Musk’s company.
The arrangement gives Reflection access to infrastructure at a time when computing availability remains one of the largest constraints on AI development. Training and serving large models require enormous quantities of advanced chips, power, and data-center capacity, creating a competitive advantage for startups that can secure infrastructure before demand rises further.
Nvidia’s backing adds another dimension. The chipmaker has a substantial interest in the development of AI models and infrastructure because the expansion of model training and inference drives demand for its accelerators.
Reflection’s launch shows that the AI race is evolving beyond the simple contest to build the largest model. Model architecture, active parameter counts, inference costs, open-weight availability, and access to computing infrastructure are becoming equally important.