Reflection AI Unveils Beam to Challenge China’s Open-Weight AI Models

Reflection AI has officially entered the increasingly competitive open-weight AI market with Beam, its first major artificial intelligence model. The US-based startup says Beam is designed primarily for coding, reasoning and agentic workloads, positioning it as a Western alternative to powerful Chinese open models from companies such as Z.ai, Alibaba and Moonshot AI. The launch comes as Chinese developers continue to dominate the open-weight segment, where downloadable models allow companies and developers to customize and operate AI systems on their own infrastructure.

Beam uses a sparse mixture-of-experts architecture containing 501 billion total parameters, while activating approximately 23 billion parameters for an individual task. This approach is intended to provide the capabilities of a much larger model without requiring the full network to be processed for every request. Reflection says the model was pretrained on 23.8 trillion tokens and underwent a large reinforcement-learning programme involving more than 100 million training rollouts on 10,500 Nvidia GB300 GPUs over four weeks.

Reflection is presenting efficiency as one of Beam’s biggest advantages. The company says the model delivers performance comparable to Z.ai’s GLM-5.2 on advanced reasoning tasks while requiring roughly three to four times less inference compute than comparable open models. It also says Beam is approaching Alibaba’s Qwen3.8-Max on coding and agentic tasks, although China’s Kimi K3 remains ahead on raw capability according to Reflection’s own evaluation. These performance comparisons are largely based on the company’s published testing and will require broader independent evaluation as the model becomes available.

The launch also has strategic significance beyond the technical competition. Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou and has received major backing from Nvidia. The company has been building toward a US alternative to Chinese open-weight systems, particularly for organizations that want customizable AI without relying on Chinese-developed technology. Reflection has also secured partnerships involving the US government and other organizations, strengthening its position in the emerging market for sovereign and controllable AI systems.

Beam is currently undergoing final red-teaming and evaluations, with early access being provided to selected users. Reflection says it plans to release the model weights, technical report, model card and developer tools later in October under an Apache 2.0 license, potentially allowing developers to inspect, modify and deploy the model more freely. If independent testing confirms Reflection’s efficiency and capability claims, Beam could become an important Western competitor in a market increasingly shaped by open AI models from China and could intensify the global race to build cheaper, customizable and high-performance AI systems.