The U.S. has a new horse in the race to build the best open-source AI model, a competition China is currently dominating, with popular models such as Kimi K3 from Moonshot and DeepSeek’s V4.
Reflection, a New York City–based AI company, today unveiled its first model, called Beam. It “advances the frontier for the Western open ecosystem,” the company told Fortune. The company said that in benchmark tests it conducted Beam achieved scores similar to those of GLM-5.2, an open model from Beijing-based Z.ai that was released back in July. That model was succeeded by GLM-5.3, which Z.ai debuted in August.
Reflection said Beam is cheaper to run than many competing models. “Through high-compute reinforcement learning, we were able to make Beam extremely efficient at reasoning—delivering competitive performance on coding and agentic tasks at a fraction of the token cost and inference time compute,” the company said. Reflection added that Beam is three to four times as efficient as rival open models from Western companies.
It is positioning Beam as a “workhorse” for enterprises, the public sector, and developers. In May, the company entered into a partnership to supply its models to the U.S. Department of Energy and the U.S. Department of War.
The company also signed a $6.3 billion compute deal with SpaceX in June for access to Nvidia GB300 chips at its Colossus 2 data center, CNBC reported. It added another $1 billion compute deal in July via cloud provider Nebius, Bloomberg reported.
Reflection was founded in March 2024 by former Google DeepMind researchers Misha Laskin, now its CEO, and Ioannis Antonoglou, the startup’s president and chief technology officer. The Nvidia-backed firm has a $25 billion valuation, Laskin told CNBC in April.
The best-known AI companies—OpenAI, Anthropic, and Google DeepMind—have primarily produced AI models that customers can only interact with through an interface, known as an API, submitting a prompt and receiving a response. By contrast, open-source models can be downloaded in their entirety by users, run on their own hardware or cloud infrastructure, and fine-tuned or modified in a variety of ways. Increasingly, large enterprises are gravitating to these open models because of the control over data security and costs that they provide, even though they lag the proprietary models in terms of performance in many cases.
“Open systems evolve faster, attract broader participation, and foster ecosystems of research and innovation that no single organization can achieve alone,” reads the company’s mission statement on its website. “They’re also safer: When systems are open to inspection, more researchers can identify vulnerabilities, improve safeguards, and contribute to their long-term resilience.”
The Western company most comparable to Reflection may be Thinking Machines Lab, founded by OpenAI’s former CTO Mira Murati. It unveiled its first model, Inkling, in July. Nvidia also makes an open model, called Nemotron, and OpenAI, Google, and Meta all offer open models as well.
Nvidia’s Nemotron model is the top-ranked, U.S.-based model on OpenRouter’s Large Language Model (LLM) leaderboard, coming in at No. 6 as of this writing. OpenRouter is a gateway that connects users to multiple routers, and it tracks which ones get the most use. When filtering the leaderboard by open models, the top ones are made by DeepSeek, Z.ai, Xiaomi, Tencent, and Nvidia. OpenAI’s gpt-oss model is the next top U.S.-based model, ranked No. 19.
Reflection is “already training the next, larger model on our road map,” the company said.
Reflection signed a $6.3 billion compute deal with SpaceX in June for access to Nvidia GB300 chips at its Colossus 2 data center, CNBC reported. It added another $1 billion compute deal in July via cloud provider Nebius, Bloomberg reported.