Reflection AI officially launched its first cutting-edge open source weight model, Beam. The Brooklyn-based startup, which has been in business for two years, claims that Beam's performance is benchmarking against mainstream open source models on advanced inference benchmarks, while at the same time drastically reducing costs. The company revealed more details in a lengthy blog post on Monday, saying that Beam is a plain text hybrid expert model, based on high computational power to enhance learning training, and has excellent capabilities in reasoning, code writing, and intelligent tasks. “Token cost and deductive computing power are only a fraction of the competition.” The total number of Beam parameters is 501 billion, and the number of active parameters is 23 billion. Pre-training uses 23.8 trillion tokens, and the context window reaches 1 million tokens. The performance claimed by Reflection has not been independently verified, but the company said that in advanced inference benchmarks, BEAM's score was comparable to Z.AI's GLM-5.2 and superior to current mainstream Western open source models, while using “3 to 4 times less” inference computing power. Reflection positions it as a “workhorse model” for enterprises, the public sector, and developers.

Reflection AI officially launched its first cutting-edge open source weight model, Beam. The Brooklyn-based startup, which has been in business for two years, claims that Beam's performance is benchmarking against mainstream open source models on advanced inference benchmarks, while at the same time drastically reducing costs. The company revealed more details in a lengthy blog post on Monday, saying that Beam is a plain text hybrid expert model, based on high computational power to enhance learning training, and has excellent capabilities in reasoning, code writing, and intelligent tasks. “Token cost and deductive computing power are only a fraction of the competition.” The total number of Beam parameters is 501 billion, and the number of active parameters is 23 billion. Pre-training uses 23.8 trillion tokens, and the context window reaches 1 million tokens. The performance claimed by Reflection has not been independently verified, but the company said that in advanced inference benchmarks, Beam's score was comparable to Z.AI's GLM-5.2, and superior to current mainstream Western open source models, while using “3 to 4 times less” inference computing power. Reflection positions it as a “workhorse model” for enterprises, the public sector, and developers.