Reflection AI officially releases Beam, which is the company's first cutting-edge open-weight AI model. This two-year-old Brooklyn-based startup claims that Beam performs on par with leading Chinese open models on advanced reasoning benchmarks, and at a much lower cost. This claim may further intensify the competition to create a "Western version of DeepSeek, Qwen, and Z.ai".

The announcement from Reflection confirmed the reports from Axios over the weekend, stating that this startup is close to releasing a new model. In a lengthy blog post published on Monday, the company revealed more details, stating that Beam is a hybrid expert model that only processes text and is trained using high-computational power reinforcement learning. It is designed to perform exceptionally well in tasks such as reasoning, programming, and agent operations, with its “token cost and inference time computational power” being just a fraction of that of its competitors.

Beam has 501 billion parameters, of which 23 billion are active parameters. It undergoes pre-training on 23.8 trillion token, and supports a context window of 1 million token. In contrast, Z.ai has a total of about 744 billion parameters in its GLM-5.2, with approximately 40 billion of those being active parameters.

The performance claims of Reflection have not yet been independently verified, but the company states that on advanced reasoning benchmarks, the score of Beam is comparable to that of Z.ai's GLM-5.2, and it outperforms today's leading Western open models while using "3 to 4 times less reasoning computing power." Reflection refers to it as the "mainstay model" for enterprises, public sectors, and developers.

Reflection positions itself as a competitor to closed laboratories such as Anthropic and OpenAI, popular open models developed by Chinese developers, as well as Western players like Mistral, Meta, and Cohere. Its most direct rival in the United States may be Inkling, which is an open model released by Thinking Machines Lab under Mira Murati in July. Reflection's own benchmark tests show that in four programming tests where both parties published results, Beam scored higher than Inkling, but Inkling is a multimodal model, while Beam only supports text.

Reflection was founded in 2024 by two former Google DeepMind researchers. According to PitchBook data, the company has raised approximately $4.7 billion from investors such as Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The pre-investment valuation of the company in its previous round of financing was $25 billion.

This startup has also been focusing on securing computing power, which is a key element in training cutting-edge models. This helps them attract customers to switch from the closed models of Anthropic and OpenAI, as well as the cheaper open-weight models from Chinese laboratories. This summer, Reflection signed agreements with SpaceX and Nebius totaling over $7 billion to ensure access to Nvidia GB300 chips by 2029.

Reflection plans to target Beam and future models at enterprises and sovereign nations. The approach is to create a " AI factory" that allows institutions to use their own proprietary data to train Reflection's AI models, thereby building customized, localized AI systems. Nvidia's CEO, Jensen Huang — whose company is also a supporter of Reflection — has long advocated for the concept of the " AI factory" and has been pushing to strengthen the open AI ecosystem; this vision will also benefit Nvidia, as these systems will be powered by their GPU.

Axios reports that hedge funds and trading companies are also among the institutions that wish to build such systems. Reflection has already begun testing the concept of cooperating with sovereign AI factories with the South Korean New World Group.

Reflection indicates that the weights and complete technical details of Beam will be released this month. At that time, they will be distributed through large-scale cloud vendors and new cloud service providers, and integrated into various open-source libraries upon release.

Reflection failed to respond in a timely manner to TechCrunch's further request for comment.