Reflection AI has introduced Beam, its first language model, in a limited preview ahead of a planned open-weight release this month. The New York startup is targeting businesses and governments that want control over their AI systems but cannot or do not want to use Chinese models.
Founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, Reflection has backing from Nvidia, Sequoia and Citigroup. The company confirmed to Semafor that its latest funding round closed in June at a $25 billion pre-money valuation, before the new investment was added.
Laskin told Semafor that sovereign-AI customers want to own and control their systems and have few suitable non-Chinese options. Reflection is positioning Beam for those buyers.
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters per token. Reflection plans to release the weights under Apache 2.0 later in October. For now, prospective users can join a waitlist for preview access.
The company’s evaluations place Beam near GLM-5.2 overall, with some coding and agentic results approaching Qwen3.8 Max. Its published DeepSWE v1.1 score is 44.4 percent, compared with 44.0 percent for GLM-5.2 and 68.0 percent for Kimi K3. Those figures are developer-reported comparisons.
Reflection also estimates that Beam uses three to four times less generation compute than models with similar performance. Its calculation uses active parameter counts and generated tokens, including reasoning. It excludes processing the initial prompt, context-dependent attention and serving overhead, so it is not a measured reduction in a customer’s bill.
Laskin said Reflection is working with the US Center for Advancing Innovation and Standards for Super Intelligence and the UK’s AI Safety Institute to assess Beam’s capabilities. He also wants developers of open models to participate in Washington’s debate over AI regulation.
That debate recently produced a White House safety accord signed by six companies. The document is a voluntary pledge, rather than a new law or an independent certification of the signers’ models.
Other Western developers are expanding their open-weight offerings. Mistral’s Large 4 preview launched on October 6, with weights planned for later in the month. Germany’s Aleph Alpha has already released Kolibri’s weights under Apache 2.0 after training on European infrastructure.
Laskin told Semafor that Reflection is already training its next model and expects it to be more capable than Beam. Beam’s immediate release plan remains the October weight download, accompanied by the technical material needed to run and evaluate it.
Sources: Reflection, Semafor
–
By the Control Plane Editorial Team