Reflection AI has officially launched Beam, its first frontier open-weight AI model, positioning the Brooklyn-based startup as a potential Western competitor to leading Chinese models like DeepSeek, Qwen, and Z.ai.
What Happened
Reflection announced Beam in a lengthy blog post on Monday, confirming earlier reports from Axios that the company was nearing a launch. The model is described as a text-only mixture-of-experts architecture trained on high-compute reinforcement learning, designed to excel at reasoning, coding, and agentic tasks. According to the company, Beam utilizes "a fraction of the token cost and inference time compute" of its rivals.
Technically, Beam features 501 billion total parameters with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and supports a 1 million token context window. For comparison, Z.ai’s GLM-5.2, a key benchmark competitor, has approximately 744 billion total parameters and 40 billion active parameters.
Reflection claims that Beam matches the performance of Z.ai’s GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open models. The company asserts that these results are achieved using 3-4x less inference compute. It is worth noting that Reflection’s performance claims have not yet been independently verified.
Why It Matters
The launch of Beam intensifies the competition for open-weight AI dominance, with Reflection explicitly targeting closed labs like Anthropic and OpenAI, as well as open model providers from China and the West, including Mistral, Meta, and Cohere. A more direct U.S. rival is Inkling, released by Mira Murati’s Thinking Machines Lab in July. While Reflection’s internal benchmarks show Beam outscoring Inkling on four coding tests, Inkling retains an advantage in multimodal capabilities, whereas Beam is text-only.
Reflection, founded in 2024 by former Google DeepMind researchers, has raised roughly $4.7 billion from backers such as Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a recent valuation of $25 billion pre-money. The company has secured significant compute resources through deals worth over $7 billion with SpaceX and Nebius to access Nvidia’s GB300 chips through 2029.
The company is pitching Beam to enterprises, the public sector, and developers as a "workhorse model" for building "AI factories." This concept involves institutions training Reflection’s models on their own proprietary data to create customized, local AI systems. Nvidia CEO Jensen Huang, whose company backs Reflection, has championed this vision to strengthen the open AI ecosystem. Hedge funds and trading firms are reportedly interested in these sovereign AI factory partnerships, with Reflection already testing the concept with Shinsegae Group in South Korea.
The Bottom Line
Reflection will release Beam’s weights and full technical details later this month, with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch. The startup did not respond to requests for additional information from TechCrunch.