Companies and governments seeking to run advanced artificial intelligence systems on their own data may soon have another option. Reflection AI, a New York City-based startup backed by Nvidia, unveiled its open-weight Beam model on 5 October 2026.
The company says Beam matches the reasoning performance of the Chinese model GLM-5.2 while using three to four times less inference compute, the processing power required when a model is put to work. Reflection also said Beam is three to four times as efficient as rival open models from Western companies.
The claims place Reflection in a contest shaped by Chinese open-source models. China dominates that competition with models including Kimi K3 from Moonshot and DeepSeek's V4.
Reflection was founded in March 2024 by Misha Laskin, its chief executive and co-founder, and Ioannis Antonoglou, its president and chief technology officer. Both had previously worked as researchers at Google DeepMind.
Beam is built for coding, reasoning and agentic work. It uses a mixture-of-experts design with 501 billion parameters overall, though only 23 billion are active at a time, and has a context window of one million tokens.
Reflection pretrained the model on 23.8 trillion tokens in under four weeks using 6,144 Nvidia GB300 GPUs. Its subsequent reinforcement learning run lasted four weeks, used 10,500 GB300 GPUs and produced more than 100 million attempts at tasks.
The startup has moved quickly to secure the computing capacity behind that work. It entered a partnership in May to supply models to the U.S. Department of Energy and the U.S. Department of War.
In June, Reflection signed a $6.3 billion compute deal with SpaceX for access to Nvidia GB300 chips at its Colossus 2 data center. It added a further $1 billion compute deal through cloud provider Nebius in July.
Laskin said the company had a $25 billion valuation in April 2026.
Reflection is aiming Beam and later models at enterprises and sovereign nations. Its proposal includes “AI factories”, which it says would allow institutions to train Reflection models on proprietary data and build customised local AI systems.
Reflection plans to release Beam's weights later in October 2026.
Comments (0)