Reflection AI, a Brooklyn-based startup, has officially unveiled Beam, its first frontier, open-weight artificial intelligence model. The company claims Beam matches the performance of leading Chinese open models on advanced reasoning benchmarks while operating at significantly lower costs, according to TechCrunch.

The new model is aimed at enterprises and sovereign nations, with Reflection AI pitching a product called "AI factories." This offering would allow institutions to build customized, local AI systems by training Reflection's models on their own proprietary data.

Reflection AI's announcement confirms earlier reporting from Axios over the weekend that the startup was preparing for a launch. The company provided further details in a blog post on Monday, October 5, describing Beam as a text-only mixture-of-experts model.

Model Specifications and Performance Claims

Beam is a 501-billion-parameter model with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and features a 1 million token context window. According to TechCrunch, Reflection AI states that Beam was trained using high-compute reinforcement learning to be effective at reasoning, coding, and agentic tasks.

The company claims Beam operates at "a fraction of the token cost and inference time compute" of its rivals. For comparison, Z.ai’s GLM-5.2, a Chinese open model, has approximately 744 billion total parameters with 40 billion active parameters.

Reflection AI's performance claims have not been independently verified, TechCrunch reported. However, the company states that on advanced reasoning benchmarks, Beam scores on par with Z.ai’s GLM-5.2 and outperforms current leading Western open models. Reflection AI also claims Beam uses "3-4x less inference compute" than its rivals.

The company describes Beam as a "workhorse model" suitable for enterprises, the public sector, and developers. In internal benchmarks cited by TechCrunch, Beam reportedly outscores Inkling, an open model from Mira Murati’s Thinking Machines Lab, on four coding tests where both models have reported results. Inkling, however, is a multimodal model, while Beam is text-only.

Company Background and Strategic Positioning

Reflection AI was founded in 2024 by two former Google DeepMind researchers. The two-year-old startup has raised approximately $4.7 billion from investors, including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook. Its most recent funding round valued the company at a $25 billion pre-money valuation.

The company has also focused on securing significant computing resources. This summer, Reflection AI signed deals collectively valued at over $7 billion with SpaceX and Nebius to gain access to Nvidia’s GB300 chips through 2029, TechCrunch reported. This compute capacity is considered crucial for training frontier models capable of attracting customers away from both closed models offered by companies like Anthropic and OpenAI, and the more affordable open-weight models from Chinese laboratories.

Reflection AI is positioning itself as a competitor against several segments of the AI market. This includes closed AI laboratories such as Anthropic and OpenAI, popular open models developed by Chinese companies, and Western players like Mistral, Meta, and Cohere. Its most direct U.S. rival in the open model space is considered to be Inkling, which was released in July.

The "AI Factory" Concept

The "AI factory" concept, which Reflection AI is promoting, involves institutions building their own customized AI systems by training Reflection’s models on their proprietary data. This vision aligns with the long-standing advocacy of Nvidia CEO Jensen Huang, whose company is a backer of Reflection AI, for strengthening the open AI ecosystem. Such a development would also benefit Nvidia, as its Graphics Processing Units (GPUs) would power these systems.

According to Axios, hedge funds and trading firms are among the entities reportedly interested in building such customized AI systems. Reflection AI has already commenced testing a sovereign AI factory partnership with Shinsegae Group in South Korea, TechCrunch reported.

Future Plans and Availability

Reflection AI plans to release Beam’s weights and full technical details later this month. The model will be distributed through hyperscalers and neoclouds, with integrations across open source libraries expected at launch, according to TechCrunch.