Reflection AI is launching Beam, its first open-weight AI model. The two-year-old startup claims Beam matches the performance of leading Chinese open models on advanced reasoning benchmarks at dramatically lower costs – a claim that could heat up the race to build a Western answer to DeepSeek, Qwen, and Z.ai.
Reflection’s announcement confirmed weekend reporting from Axios that the launch was imminent, with new details shared in a blog post Monday. Beam is a text-only mixture-of-experts model trained on high-compute reinforcement learning to handle reasoning, coding, and agentic tasks at a fraction of the token cost and inference time compute of rivals.
Featuring a robust architecture with 501 billion total parameters–23 billion of which are active–Beam has been pre-trained on an impressive 23.8 trillion tokens and offers a context window of 1 million tokens. In comparison, Z.ai’s GLM 5.2 has about 744 billion parameters, with 40 billion actively utilized.
Although Reflection’s assertions about Beam’s performance have not been independently verified, the company claims that in terms of advanced reasoning benchmarks, Beam performs similarly to Z.ai’s GLM-5.2 while using 3-4 times less inference compute than leading Western open models. This positions Beam as a valuable resource for enterprises, public sectors, and developers.
Reflection is strategically placing itself in competition not only with closed labs like Anthropic and OpenAI but also with popular open models from Chinese developers and established Western companies like Mistral, Meta, and Cohere. Its nearest U.S. competitor is Inkling, an open model from Mira Murati’s Thinking Machines Lab, which was released in July. Reflection’s benchmarks indicate that Beam outperforms Inkling on four coding tests where both reported results, although it should be noted that Inkling is a multimodal model while Beam is strictly text-based.
Founded in 2024 by former researchers from Google DeepMind, Reflection AI has raised approximately $4.7 billion from notable investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The latest funding round established the company’s pre-money valuation at $25 billion. This significant financial backing has enabled Reflection to secure vital compute resources necessary for training advanced models that can draw customers away from Anthropic and OpenAI’s closed systems, as well as from less expensive open-weight models from Chinese labs.
This summer, Reflection entered into agreements totaling over $7 billion with firms such as SpaceX and Nebius to gain access to Nvidia’s GB300 chips through 2029. These partnerships are crucial as Reflection seeks to develop Beam and future models aimed at enterprises and sovereign nations. The vision is to create “AI factories,” which would allow institutions to construct customized AI systems using their proprietary data.
Axios reported that hedge funds and trading firms are among those eager to build such systems. Reflection has already initiated a collaboration with Shinsegae Group in South Korea to explore the idea of a sovereign AI factory. Regarding Beam, Reflection intends to release its weights and comprehensive technical details this month, distributing through hyperscalers and neoclouds while ensuring integrations across open-source libraries at launch.
