Reflection AI, a two-year-old startup based in Brooklyn, announced the debut of Beam, its first frontier open-weight model. The company positions Beam as a direct Western answer to Chinese offerings such as DeepSeek, Qwen and Z.ai, emphasizing lower compute costs. The announcement follows earlier reporting that the firm was nearing a launch, and it was detailed in a lengthy blog post released on Monday.
Beam comprises a total of 501 billion parameters, of which 23 billion are active during inference. It was pretrained on roughly 23.8 trillion tokens and supports a context window of one million tokens. For comparison, Z.ai’s GLM-5.2 model has about 744 billion total parameters and 40 billion active parameters, highlighting Beam’s more compact active footprint.
Reflection claims that Beam achieves parity with Z.ai’s GLM-5.2 on advanced reasoning benchmarks while delivering three to four times lower inference compute. The company also states that Beam outperforms leading Western open models on the same tests. These performance assertions have not been independently verified, and the benchmarks are presented only in Reflection’s internal results.
The startup frames Beam as a competitor to closed-lab offerings from Anthropic and OpenAI, as well as open models from Mistral, Meta and Cohere. Its closest U.S. rival appears to be Inkling, an open model from Thinking Machines Lab released in July. Reflection’s internal tests show Beam surpassing Inkling on four coding assessments, though Inkling provides multimodal capabilities while Beam remains text-only.
Founded in 2024 by two former Google DeepMind researchers, Reflection has raised approximately $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook. The latest financing valued the company at a $25 billion pre-money valuation. To secure the compute needed for frontier training, Reflection signed contracts worth more than $7 billion with SpaceX and Nebius for access to Nvidia’s GB300 chips through 2029.
Reflection markets Beam and future models toward enterprises and sovereign nations, promoting the concept of “AI factories” that let customers train the models on proprietary data. The company has begun piloting a sovereign AI-factory partnership with South Korea’s Shinsegae Group. Nvidia’s CEO Jensen Huang, a backer of Reflection, has long advocated for the AI-factory model, which would also expand demand for Nvidia GPUs.
Reflection says it will make Beam’s weights and full technical documentation publicly available later this month. Distribution is planned through major hyperscalers and emerging “neocloud” providers, with integrations across open-source libraries at launch. The open-weight release aims to accelerate adoption by developers, public-sector users and corporate clients seeking a cost-effective alternative to closed AI services.