Nvidia-backed Reflection AI has unveiled Beam, its first open-weight artificial intelligence model, as the US startup looks to compete with Chinese AI developers that have gained ground with lower-cost and customisable models.
Beam is a text-based mixture-of-experts model built for reasoning, coding and agentic workloads. It has 501 billion total parameters but activates 23 billion parameters for each token, an architecture designed to reduce the computing resources required during inference.
Reflection AI said Beam is competitive with Z.ai's GLM-5.2 across several reasoning and coding evaluations and is approaching Alibaba's Qwen3.8-Max on some tasks. The company has positioned efficiency as a key differentiator, claiming the model can deliver comparable reasoning performance to GLM-5.2 while using three to four times less inference compute. These benchmark results have been reported by Reflection and have not yet been independently validated.
The model has been trained on 23.8 trillion tokens and uses a sparse architecture that activates only a portion of its overall network for each task. Beam is designed particularly for software development, multi-step reasoning and AI agents that use tools to complete tasks.
Reflection plans to release Beam's model weights, technical report, model card and associated tooling later in October under the Apache 2.0 licence. The model is currently available through an early-access programme while the company completes evaluations and safety testing.
The open-weight approach will allow developers and organisations to run and customise the model on their own infrastructure once the weights are released. This differs from proprietary models where access is generally provided through applications or APIs and the underlying model weights remain closed.
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company has focused on developing AI systems for software engineering and autonomous coding workflows.
The startup is backed by Nvidia and other investors and has been building the computing infrastructure required to train increasingly large models. Earlier this year, Reflection also secured additional computing capacity through an agreement involving SpaceX's Colossus 2 data centre.
Beam arrives as Chinese developers including DeepSeek, Alibaba, Moonshot AI and Z.ai continue to expand their open-model portfolios. Reflection acknowledges that models such as Moonshot AI's Kimi K3 remain ahead on raw capability in some evaluations.
The company is already working on a larger successor to Beam, signalling that the newly released model is the first step in its broader push to develop competitive open-weight AI systems.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.