Reflection, a New York-based artificial intelligence company, has unveiled Beam, an open-source model it says is designed to strengthen the Western challenge to China’s increasingly influential AI sector.
The company said Beam achieved results comparable with GLM-5.2, an open model released by Beijing-based Z.ai in July, in benchmark testing conducted by Reflection. Z.ai subsequently launched GLM-5.3 in August.
Reflection is presenting Beam as a practical “workhorse” for businesses, public-sector organisations and software developers. It says the model is particularly suited to coding and agent-based tasks, in which AI systems carry out a sequence of actions on a user’s behalf.
The company said Beam’s use of high-compute reinforcement learning had made it more efficient at reasoning, reducing the amount of computing and processing time needed to produce results.
“Through high-compute reinforcement learning, we were able to make Beam extremely efficient at reasoning—delivering competitive performance on coding and agentic tasks at a fraction of the token cost and inference time compute,” Reflection said.
Reflection also claims Beam is three to four times more efficient than open models produced by rival Western companies. The performance and efficiency figures have been provided by the company and were based on its own testing.
Reflection’s open-source push
Unlike models that are accessed mainly through an online interface, open-source systems can be downloaded, operated on a customer’s own hardware or cloud infrastructure, and adapted or fine-tuned for different uses.
That flexibility has made such models increasingly attractive to large companies seeking greater control over data security and running costs, although open systems can still trail proprietary models in performance.
Reflection was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Laskin is the company’s chief executive, while Antonoglou is its president and chief technology officer.
The company has entered into a partnership to supply models to the US Department of Energy and the US Department of War. It has also agreed a $6.3 billion computing deal with SpaceX for access to Nvidia GB300 chips at the Colossus 2 data centre, and a further $1 billion arrangement through cloud provider Nebius.
Laskin said in April that Nvidia-backed Reflection was valued at $25 billion.
The launch comes as Chinese developers dominate several rankings of widely used open models. Models from Moonshot, DeepSeek, Z.ai, Xiaomi and Tencent feature prominently, while Nvidia’s Nemotron is the highest-ranked US-based model on OpenRouter’s large language model leaderboard, at number six.
OpenAI’s gpt-oss is the next highest-ranked US model on the same leaderboard, at number 19. Other Western companies offering open models include Google, Meta and Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati.
Reflection said it was already training a larger model as part of its development plans.
