Reflection AI has unveiled Beam, its first open-weight model, claiming it matches the performance of top Chinese rivals while costing significantly less to run.
The Brooklyn-based startup, which has existed for two years, says Beam delivers comparable results on advanced reasoning benchmarks to models like DeepSeek, Qwen, and Z.ai but at a fraction of the compute expense. This claim could intensify the push to build a Western alternative to the current leaders in open-source AI.
Details released on Monday confirm reports from Axios that the company was nearing a launch. Beam is described as a text-only mixture-of-experts model trained using high-compute reinforcement learning. The goal is effectiveness in reasoning, coding, and agentic tasks without the heavy token costs or inference times seen in competing systems.
The model contains 501 billion total parameters, with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and supports a context window of one million tokens. By comparison, Z.ai’s GLM-5.2 has approximately 744 billion total parameters and 40 billion active parameters.
While no independent verification exists yet for these performance claims, Reflection states Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning tests. The company asserts it outperforms current leading Western open models while using three to four times less inference compute. Reflection markets the model as a workhorse for enterprises, the public sector, and developers.
The startup positions itself against closed labs such as Anthropic and OpenAI, alongside popular open models from Chinese developers. It also competes with Western players like Mistral, Meta, and Cohere. A direct US rival might be Inkling, an open model from Mira Murati’s Thinking Machines Lab released in July. Reflection’s own benchmarks show Beam outscoring Inkling on four coding tests where both reported results, though Inkling is multimodal and Beam is text-only.
Reflection was founded in 2024 by two former Google DeepMind researchers. Backers include Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook. The company raised roughly $4.7 billion, with its last round valuing it at $25 billion pre-money.
The startup has secured access to Nvidia’s GB300 chips through 2029 via deals worth more than $7 billion with SpaceX and Nebius. This compute access is vital for training frontier models that could draw customers away from closed systems by Anthropic and OpenAI, as well as cheaper open-weight options from Chinese labs.
Beam and future models target enterprises and sovereign nations. The strategy involves building “AI factories,” allowing institutions to train Reflection’s models on their own proprietary data to create local, customised systems. Jensen Huang, CEO of Nvidia, has long championed this concept. His company backs Reflection, and the vision benefits Nvidia as its GPUs power these local systems.
Hedge funds and trading firms are reportedly eager to build such systems. Reflection has already begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea.
Weights and full technical details for Beam will be released this month. Distribution will occur through hyperscalers and neoclouds, with integrations across open source libraries available at launch.
Reflection did not respond to requests for further information.
What it means
For teams building internal tools, this announcement suggests a potential shift away from expensive, closed APIs. If the claims hold true, organisations can run sophisticated reasoning and coding tasks locally without paying per-token fees or relying on external cloud providers. The focus on “AI factories” implies a move toward on-premise control, where businesses own the compute and the data, reducing dependency on major tech platforms.
