TLDR Tech / Launches
New open AI model Beam offers 1M token context, fast inference
Reflection and Hugging Face have unveiled Beam, a large open-source language model with 501 billion parameters. This Mixture-of-Experts model boasts a context window of one million tokens and includes a custom harness built in Rust. It is compatible from day one with transformers, vLLM, and SGLang. Optimized for coding and agentic applications, Beam employs grouped-query attention and a deepseek-v3 style expert routing system to achieve inference speeds comparable to a 23 billion parameter model. The model is available under the Apache 2.0 license, with checkpoints in BF16, FP8, and NVFP4 formats on Hugging Face.
Read the full story