Reflection’s Beam pairs 501 billion parameters with a leaner compute claim

What Reflection says Beam can do

Beam is a mixture-of-experts model with 23 billion parameters active at once. Reflection says it was pretrained on 23.8 trillion tokens, has a one-million-token context window and was trained with reinforcement learning to handle reasoning, coding and AI-agent tasks.

The comparison point is Z.ai’s GLM-5.2, which has about 744 billion parameters in total and 40 billion active. Reflection says Beam matches it on difficult reasoning benchmarks and outperforms leading Western open models while requiring three to four times less inference compute. The company describes Beam as a production model for businesses, government agencies and developers.

That evidence has limits. Reflection’s benchmark claims remain unverified independently, and the company has published results for only four coding tests that allow a direct comparison with Inkling, the open model from Mira Murati’s Thinking Machines Lab released in July. On those tests, Reflection says Beam leads. But Inkling can handle multiple data types, while Beam accepts text only.

The competitive field also includes closed labs such as Anthropic and OpenAI, Chinese open-model developers, and Western companies Mistral, Meta and Cohere. A benchmark lead is useful; it does not by itself establish that Beam is the more capable or more useful system.

Compute is part of the product

Reflection was founded in 2024 by two former Google DeepMind researchers. PitchBook says it has raised about $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners. Its latest funding round valued the company at $25 billion before new investment.

The company has also lined up access to the hardware needed to train and serve advanced models. This summer, it signed agreements with SpaceX and Nebius worth more than $7 billion in total, securing access to Nvidia GB300 chips through 2029.

The compute claim matters because Reflection is selling more than a model. It is developing “AI factories” for companies and sovereign states: customers would train Reflection models on their own data and run a locally tailored AI system. Nvidia CEO Jensen Huang, whose company invested in Reflection, has long promoted AI factories and an open AI ecosystem. Nvidia also benefits if those systems run on its GPUs.

Axios reported interest from hedge funds and trading firms. Reflection has begun exploring a partnership with South Korea’s Shinsegae Group to build a sovereign AI factory. Beam is expected to be available through hyperscalers and newer cloud providers, as well as integrated with open-source libraries.

The test is whether the pitch travels

I think the more consequential question is not whether Beam can edge out Inkling on four coding tests. It is whether customers can use the promised efficiency in the settings Reflection is targeting: their own data, local systems and demanding workloads. The announcement offers a compute comparison, but no independently checked results or evidence yet that the proposed AI-factory model works at scale.

Reflection plans to release Beam’s weights and full technical documentation this month. That will give researchers and developers a chance to inspect the model and test the claims. Until then, the company’s $7 billion-plus hardware agreements make its ambitions tangible, while the central promise—strong performance at substantially lower inference cost—remains Reflection’s own.