Nvidia-Backed Reflection AI Launches Beam to Challenge Chinese Open-Weight AI Models
Updated: 3 minutes ago
Nvidia-backed Reflection AI has launched Beam, its first open-weight artificial intelligence model, as the startup looks to compete with increasingly capable Chinese models from companies including Z.ai, Alibaba, DeepSeek and Moonshot AI.
Beam is designed primarily for coding, reasoning and agentic AI workloads, where AI systems can plan and carry out tasks with greater autonomy. Reflection says the model is competitive with Z.ai’s GLM-5.2 and is approaching Alibaba’s Qwen3.8-Max in coding and agentic performance.
The model uses a Mixture-of-Experts architecture with 501 billion total parameters, but activates only 23 billion parameters for each token. This approach is intended to reduce the computing resources required to run the model while maintaining strong performance.
Reflection claims Beam can achieve results comparable to GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. The comparison is based on Reflection’s own estimates rather than measured serving costs, so independent testing will be needed to establish how those efficiency gains translate into real-world deployments.
Taking on Chinese open-weight AI models
Chinese AI developers have become particularly competitive in open-weight models, which give developers and businesses greater flexibility to download, customise and deploy AI systems on their own infrastructure. Models from DeepSeek, Alibaba’s Qwen, Z.ai and Moonshot AI have helped make high-performance AI available at lower costs, increasing pressure on other developers to produce similarly capable alternatives.
Reflection is positioning Beam directly in this market rather than claiming to outperform proprietary frontier systems across every category. Its own benchmark results show a mixed picture, with Beam competitive against GLM-5.2 in several areas while newer Chinese models remain ahead on some tests. Reflection acknowledges that Moonshot AI’s Kimi K3 remains stronger in raw capability, arguing that Beam’s main advantage is inference efficiency.
Built with Nvidia computing power
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou and focuses heavily on AI systems capable of automating software development. Beam was pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 GPUs. Reflection says its reinforcement-learning programme then generated more than 100 million rollouts using 10,500 GB300 GPUs over four weeks.
Beam is currently undergoing final red-team testing and evaluations, with early access available to selected users. Reflection plans to release the model weights under an Apache 2.0 licence, together with its technical report, model card and developer resources later in October.
The launch adds another competitor to the fast-moving open-weight AI market, where Chinese models have emerged as some of the strongest alternatives to proprietary systems.
