Reflection AI announced Beam, its first frontier open-weight model, positioning the system as a cost-efficient alternative to some leading Chinese and Western open models. The company describes Beam as a text-only mixture-of-experts model trained with high-compute reinforcement learning and aimed at reasoning, coding and agentic tasks.
Model design and benchmark claims
Beam is reported to have 501 billion total parameters with 23 billion active parameters and was pretrained on 23.8 trillion tokens. The model offers a 1 million token context window. Reflection compares Beam to Z.ai’s GLM-5.2 (roughly 744 billion total parameters and 40 billion active) and says Beam scores on par with GLM-5.2 on advanced reasoning benchmarks.
The company also claims Beam outperforms current Western open models while using “3-4x less inference compute.” Reflection describes the release as a “workhorse model” for enterprises, the public sector and developers. Those performance and efficiency claims have not been independently verified; they are reported by the company.
Competitive positioning and test results
Reflection positions Beam against closed labs such as Anthropic and OpenAI, Chinese open models, and Western players including Mistral, Meta and Cohere. The company notes a close U.S. peer is Inkling, the open model from Thinking Machines Lab released in July. Reflection’s internal benchmarks indicate Beam outscores Inkling on four coding tests where both report results, while noting that Inkling is multimodal and Beam is text-only.
Reflection was founded in 2024 by two former Google DeepMind researchers. According to PitchBook figures cited by the company, it has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, with a last-round pre-money valuation of $25 billion.
Reflection has also secured compute capacity through deals exceeding $7 billion with SpaceX and Nebius to access Nvidia GB300 chips through 2029. The company is marketing an “AI factories” concept to help enterprises and sovereign nations train customized local systems on proprietary data, and has begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea.
Reflection says it will publish Beam’s weights and full technical details this month, distributing them through hyperscalers and neoclouds and integrating with open-source libraries at launch.
Original source: TechCrunch AI