Brooklyn-based AI startup Reflection AI has unveiled Beam, its first open-weight “frontier” model. The two-year-old company claims that Beam delivers the same performance as leading Chinese open models on advanced reasoning tests at a much lower cost. These claims have not been independently verified.
According to the company’s blog post, Beam is a text-only “mixture-of-experts” model trained to be effective at reasoning, coding, and agentic tasks with “a fraction of the token cost and inference compute” of its rivals. The model’s technical specifications:
- 501 billion total parameters, 23 billion active parameters
- Pre-training on 23.8 trillion tokens
- 1 million token context window
For comparison, Z.ai’s GLM-5.2 model has approximately 744 billion total and 40 billion active parameters. Reflection claims that Beam achieves the same scores as GLM-5.2 while outperforming leading Western open models with “3-4 times less inference compute.”
Reflection positions itself against closed labs such as Anthropic and OpenAI, Chinese open models, and Western players like Mistral, Meta, and Cohere. Its most direct US rival is Inkling, the model released by Mira Murati’s Thinking Machines Lab in July. In Reflection’s tests, Beam beats Inkling on four shared coding benchmarks; however, Inkling is multimodal while Beam is text-only.
Founded in 2024 by two former Google DeepMind researchers, Reflection has raised approximately $4.7 billion from investors such as Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to PitchBook data; the company was given a $25 billion pre-money valuation in its latest round. This summer, the startup signed deals with SpaceX and Nebius totaling more than $7 billion, securing access to Nvidia GB300 chips through 2029.
Reflection markets Beam to enterprises and sovereign states, promoting the idea of “AI factories” that would allow organizations to build local AI systems with their own data. The company has begun testing this concept in South Korea with Shinsegae Group. Beam’s weights and technical details will be published this month.
Background
Beam is not a new name in the FikirPilot archive: we have published 2 stories mentioning this name in the last 90 days; the most recent is dated October 4, 2026.
Term: inference
Inference is a trained model generating a response to a new input; unlike training, it incurs a new cost with each use.