Reflection AI officially announces Beam, its first frontier open-weight AI model. The two-year-old Brooklyn-based startup claims Beam can match the performance of leading Chinese open models on advanced inference benchmarks at dramatically lower cost, a claim that could intensify the race to build a Western answer to DeepSeek, Qwen and Z.ai.
Reflection’s announcement confirms a report from Axios over the weekend that the startup was nearing launch. The company shared new details in a lengthy blog post on Monday, in which it described Beam as a text-only expert mixture model trained with high-computing reinforcement learning to effectively perform inference, coding, and agent tasks at “a fraction of the token cost and inference time calculations” of its competitors.
Beam is a 501 billion parameter model with 23 billion active parameters. It is pre-trained with 23.8 trillion tokens and has a context window of 1 million tokens. By comparison, Z.ai’s GLM-5.2 has approximately 744 billion total parameters, of which 40 billion are active.
Reflection’s performance claims have not been independently verified, but on advanced inference benchmarks, Beam scores on par with Z.ai’s GLM-5.2 and outperforms today’s leading Western open models while using “three to four times less inference computing,” the company says. Reflection calls this the “go-to model” for businesses, the public sector, and developers.
Reflection positions itself against closed labs like Anthropic and OpenAI, popular open models from Chinese developers, and Western players like Mistral, Meta, and Cohere. Its most direct rival in the US may be Inkling, Mira Murati’s Thinking Machines Lab’s open model released in July. Reflection’s own benchmarks show that Beam outperforms Inkling in four coding tests, both of which have reported results, but Inkling is a multimodal model and Beam is text-only.
Reflection was founded in 2024 by two former Google DeepMind researchers and has raised about $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook. The previous round valued the company at a pre-money valuation of $25 billion.
The startup is also locking up compute, a key element needed to train frontier models that could lure customers away from closed models from Anthropic and OpenAI or cheaper open-weight models from Chinese labs. This summer, Reflection signed a deal worth more than $7 billion combined with SpaceX and Nebius to secure access to Nvidia’s GB300 chips through 2029.
Reflection aims to be a beam and future model in corporations and sovereign nations. This product enables agencies to build their own customized local AI systems by training Reflection’s AI models on their own data. Nvidia CEO Jensen Huang, who is behind Reflection, has long championed the idea of an “AI factory” and pushed for a stronger open AI ecosystem. This vision is also a win-win for Nvidia, where GPUs power systems.
Axios reported that hedge funds and trading companies are also keen to build such systems. Reflection has already started testing the concept of a sovereign AI factory partnership with South Korea’s Shinsegae Group.
Reflection says it will release weights and full technical details for Beam this month, with distribution via hyperscalers and neoclouds and integration between open source libraries at launch.
Reflection did not respond to TechCrunch’s request for more information.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
