Reflection launches Beam open-weight AI model

Free Instant SEO Audit

Glowing processor represents the Beam open-weight AI model shattering a compute barrier

By NewSunSEO October 15, 2026

Reflection has officially launched Beam, its first frontier open-weight AI model, claiming performance on par with leading Chinese open models on advanced reasoning benchmarks while using far less inference compute. The 501-billion-parameter mixture-of-experts model is positioned as a Western answer to Chinese open-weight systems and a lower-cost option for enterprises, developers, and public-sector buyers who want a model they can run and customize themselves.

What Beam is and how it was built

Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and supports a one-million-token context window. Reflection trained it with high-compute reinforcement learning to specialize in reasoning, coding, and agentic tasks, then released the weights and full technical details this month. Distribution runs through hyperscalers and neoclouds, with integrations across open-source libraries at launch.

For comparison, Z.ai’s GLM-5.2 has roughly 744 billion total parameters and 40 billion active, which is the scale Beam is measured against in Reflection’s own benchmarks.

What the benchmarks show, and what they do not

Reflection says Beam scores on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and outperforms today’s leading Western open-weight models while using 3-4x less inference compute. The company describes Beam as a workhorse model aimed at enterprise workloads, the public sector, and developers who need a controllable, locally tunable system.

Reflection’s benchmarks also show Beam outscoring Inkling, the open model from Thinking Machines Lab, on four coding tests where both models report results. Inkling is multimodal, so the two systems are not directly comparable across every task. None of these claims have been independently verified, which is worth keeping in mind as the weights and full technical report are reviewed in the open.

How Beam fits into the open-weight landscape

The launch lands in a market that is now crowded on both sides. Chinese labs have shipped widely adopted open-weight models that have pulled developer mindshare away from closed U.S. labs. Reflection is positioning Beam as a Western counterpart that enterprises, sovereign customers, and regulated industries can adopt without the data-residency or supply-chain concerns that come with foreign models. The competitive set is broad: closed labs like Anthropic and OpenAI, Western open-weight players like Mistral, Meta, and Cohere, and a growing list of Chinese open-weight releases.

What the company is building around Beam

Reflection is aiming Beam and future models at enterprises and sovereign nations, with a product it calls an AI factory. The idea is to let institutions train Reflection’s base models on their own proprietary data and run the resulting system locally, rather than sending sensitive data to a third-party API. Reflection has begun testing the concept through a partnership with Shinsegae Group in South Korea, and reports indicate hedge funds and trading firms are among the early interested buyers.

The strategy rests on a compute footprint Reflection has been locking in over the past year. This summer the company signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029, which is the kind of long-dated compute commitment frontier training demands.

Funding and backers

Reflection was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion to date, with backing from Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The last funding round valued the company at $25 billion pre-money, a level that reflects the premium the market is placing on credible Western open-weight contenders at a time when Chinese open models have set the performance bar that closed Western labs are now trying to clear at lower cost.

FAQ

What is Beam and who built it?

Beam is an open-weight AI model from Reflection, a Brooklyn-based startup founded in 2024 by former Google DeepMind researchers. It is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, pretrained on 23.8 trillion tokens with a one-million-token context window.

How does Beam compare to Chinese open-weight models?

Reflection says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open-weight models while using 3-4x less inference compute. The claims are not independently verified, and the full technical report is being released alongside the weights this month.

What is Reflection’s AI factory concept?

An AI factory is a product offering that lets enterprises and sovereign nations train Reflection’s models on their own proprietary data and run the resulting system locally. Reflection is testing the idea through a partnership with Shinsegae Group in South Korea.