Reflection AI has officially unveiled Beam, its first advanced AI model with open weights. The Brooklyn-based company claims that Beam is on par with leading Chinese open models on complex reasoning tests, but requires significantly less computational effort. This could significantly intensify the race to create a Western alternative to DeepSeek, Qwen, and Z.ai.
Reflection AI has officially unveiled Beam, its first advanced AI model with open weights. The Brooklyn-based company claims that Beam is on par with leading Chinese open models on complex reasoning tests, but requires significantly less computational effort. This could significantly intensify the race to create a Western alternative to DeepSeek, Qwen, and Z.ai.
According to Reflection’s official blog, Beam is a text-based model based on the Mixture-of-Experts (MoE) architecture. It is trained using intensive reinforcement learning and is designed to efficiently solve reasoning, coding, and autonomous agent tasks. At the same time, the model spends “only a fraction of the token cost and computing time to infer” compared to competitors, writes TechCrunch.
Parameters: Beam has a total of 501 billion parameters, of which 23 billion are active. For comparison, Z.ai’s GLM-5.2 model has about 744 billion parameters (40 billion active).
Context window: 1 million tokens.
Previous training: conducted on 23.8 trillion tokens.
While Reflection’s claims have not yet been independently verified, the company claims that in reasoning benchmarks, Beam matches Z.ai GLM-5.2 and outperforms current leading Western open-source models, using "3-4 times less computing power during inference."
The creators position it as a "workhorse" for the corporate sector, government agencies, and developers.
Reflection contrasts itself with both closed labs (OpenAI, Anthropic) and open developers from China and the West (Mistral, Meta, Cohere).
Beam’s direct American competitor is the Inkling model from Thinking Machines Lab (founded by Mira Murati), released in July. According to Reflection’s internal tests, Beam outperforms Inkling in four coding benchmarks, even though Inkling is a multimodal model, while Beam works exclusively with text.
Reflection is targeting Beam at corporate and government organizations. Its main offering is the concept of “AI factories,” which allow organizations to build their own local systems by training Reflection models on their own private data.
Reflection has already begun testing the concept of a sovereign “AI factory” in a partnership with the South Korean company Shinsegae Group.
Beam weights and full technical details are scheduled to be released this month. The model will be available through major cloud providers, neo-clouds, and integrated into open libraries.
The startup Reflection was founded in 2024 by two former Google DeepMind researchers. According to PitchBook, the company has already raised about $4.7 billion from Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and its valuation has reached $25 billion (pre-money).
The startup is actively purchasing computing resources to train new-generation models. Over the summer, Reflection signed deals with SpaceX and Nebius worth a total of over $7 billion, securing access to the latest Nvidia GB300 chips until 2029.
