Reflection releases Beam, a 501B open-weight model built for coding agents

Reflection, a new US AI lab, has published the weights for Beam, its first model, on Hugging Face under the Apache 2.0 license. Hugging Face announced the release on October 16, 2026, after Reflection first previewed the model earlier in the month.

Beam is a mixture-of-experts model with about 501 billion parameters in total, but only around 23 billion are used for each token, so it runs closer to the speed of a mid-sized model. It reads and writes text only, handles up to 1 million tokens of context, and is aimed mainly at coding and agent work such as fixing GitHub issues, running terminal tasks and calling tools through MCP. Reflection's own numbers put it at about 81 percent on SWE-bench Verified and 80 percent on Terminal Bench 2.1.

The weights come in BF16, FP8 and NVFP4 versions. The 4-bit NVFP4 file is roughly 280 GB, and Hugging Face suggests four Nvidia B200 GPUs for comfortable serving. Beam works on day one with transformers, vLLM and SGLang, ships with a Rust-based agent harness, and is also offered through hosted providers including Baseten, Fireworks and Together. Reflection is offering $10,000 in Hub credits for the best projects built on it.

Why it matters

A permissively licensed, frontier-scale coding model gives companies another option they can host and fine-tune themselves instead of renting a closed API.

Source: Hugging Face