TLDR Tech / Launches
Nvidia-backed startup Reflection AI releases Beam model
Reflection AI, a startup supported by Nvidia, has unveiled Beam, its initial open-weight model. This mixture-of-experts model has 501 billion parameters and is designed for tasks involving coding, reasoning, and agentic functions. Beam was trained on 23.8 trillion tokens and features a context window of one million tokens. It is slated for release under an Apache 2.0 license this month. According to internal benchmarks, Beam demonstrates comparable performance to GLM-5.2 in advanced reasoning while requiring substantially less computational power for inference.
Read the full story