Reflection has unveiled Beam, its first open-weight model designed for coding, logical reasoning, and agentic tasks. The mixture-of-experts model features 501 billion total parameters, with 23 billion active per token. According to the company, Beam matches GLM 5.2 performance while utilizing three to four times less compute, positioning it as a competitor to models from Deepseek and Qwen.
The training process involved a reinforcement learning phase using 10,500 Nvidia GB300 GPUs over four weeks. Reflection reported emergent capabilities, noting that the model independently improved its web browsing skills during training. Beam includes a tunable parameter that allows users to adjust reasoning depth, balancing compute costs against output quality.
Safety alignment was achieved by merging Beam with a secondary model trained on specific guidelines. Reflection plans to release a technical report, documentation, and model weights under the Apache 2.0 license later this month. Currently, an early version is available to select users while final safety testing continues.
Founded in 2024 by former Google Deepmind researchers Misha Laskin and Ioannis Antonoglou, Reflection raised $130 million in seed funding in March 2025. The company later secured $2 billion in October 2025 at an $8 billion valuation, with participation from Nvidia. Reflection has also finalized billion-dollar compute agreements with SpaceX and Nebius.
