Reflection AI unveils Beam, its first frontier open-weight model

Beam has 501 billion parameters, and Reflection AI says it matches leading Chinese open models on advanced reasoning benchmarks. The company describes the release as its first frontier open-weight model, intended for reasoning, coding and agentic tasks. Its performance claims have not been independently verified, leaving independent assessment central to any comparison.

Beam is a text-only mixture-of-experts model, a design that activates only part of its total parameters for a given task. Reflection AI says 23 billion parameters are active, and that pretraining used 23.8 trillion tokens. The model also has a 1 million token context window, which describes how much text it can process at once.

Beam targets institutional deployments

Reflection AI presents Beam as a workhorse for enterprises, the public sector and developers. The company’s broader pitch is that institutions could use its models to build customised local AI systems trained on their own proprietary data. That proposition makes model access and deployment choices relevant to organisations handling sensitive information.

The company calls this approach AI factories, a term for systems that allow institutions to adapt models using their own data. Reflection AI has said it is aiming Beam and future models at enterprises and sovereign nations. The available account does not set out deployment terms, access conditions or the safeguards that would govern customer data.

Reflection AI reports lower inference needs

Reflection AI says Beam performs on par with Z.ai GLM-5.2 on advanced reasoning benchmarks and surpasses leading Western open models. It also claims Beam uses 3-4x less inference compute, meaning computational resources used to generate answers after a model has been trained. No independent verification of these comparisons is reported.

Those claims matter to buyers comparing the operational demands of open-weight systems. However, benchmark results alone do not establish how a model will perform across an organisation’s tasks or what its full deployment costs will be. The source material does not provide independent test results or a broader account of operating costs.

Funding supports a costly training effort

Reflection AI was founded in 2024 by two former Google DeepMind researchers, according to the source account. It has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, per PitchBook. Its last funding round valued the company at a $25 billion pre-money valuation, which is the company value before new investment is added.

The company has also arranged access to computing hardware. The source account says Reflection AI signed deals with SpaceX and Nebius, collectively worth more than $7 billion, for access to Nvidia GB300 chips through 2029. That reported infrastructure commitment sits alongside the stated aim of training frontier models and serving institutional customers.

The performance figures remain company claims, and the source reports no independent benchmark confirmation. The release account also leaves deployment conditions and customer access terms unspecified. Independent results would clarify whether the comparisons hold across relevant reasoning tasks. Further details about availability and institutional safeguards would help buyers assess adoption.