Reflection AI introduced Beam, a text-only model with 23 billion active parameters, and says it matches leading Chinese open models using less inference compute.
Reflection AI unveiled Beam on October 5, 2026, presenting it as the Brooklyn startup’s first frontier open-weight model. The text-only mixture-of-experts model has 501 billion total parameters, of which 23 billion are active, and a one-million-token context window. Reflection says it was pretrained on 23.8 trillion tokens and trained with reinforcement learning for reasoning, coding, and agent tasks.
The company claims Beam performs on par with Z.ai’s GLM-5.2 on advanced reasoning benchmarks and exceeds leading Western open models while using three to four times less inference compute. Those performance comparisons have not been independently verified, according to TechCrunch. The article also notes that Beam is text-only, while one of the models it is compared with, Thinking Machines Lab’s Inkling, supports multiple modalities.
Reflection is targeting enterprises, public-sector organizations, and developers. Its longer-term commercial pitch involves “AI factories”: customized local AI systems that institutions could build by training Reflection models on proprietary data. The company has begun testing that approach in a sovereign AI factory partnership with South Korea’s Shinsegae Group, TechCrunch reported.
The model launch comes as Reflection builds access to computing capacity. The company has raised roughly $4.7 billion, according to PitchBook, and its last funding round valued it at a $25 billion pre-money valuation. TechCrunch also reported that Reflection signed agreements with SpaceX and Nebius worth more than $7 billion collectively to secure access to Nvidia GB300 chips through 2029.
Reflection said it plans to release Beam’s weights and full technical details in October, with distribution through hyperscalers and neocloud providers and integrations across open-source libraries at launch. Until those materials are available and the company’s benchmark claims can be examined independently, the reported performance and efficiency figures remain Reflection’s claims rather than independently established results.