Welcome Beam: a 501B open model you can own, customize, and scale

Hugging Face has added Beam, the first model from the new US lab Reflection. It’s a big one: a 501B-parameter mixture-of-experts model with 23B active parameters per token, text input only, a 1M-token context window, and an Apache 2.0 license. So yes, it’s open, and not in the vaguely “open-ish” way that usually comes with a catch or a very long wait.

Beam is aimed at coding and agentic workloads, which is code for “let teams point it at real work and mess with it until it fits.” Reflection and Hugging Face are also pushing the customization angle hard. The model comes with recipes for inference, supervised fine-tuning, and reinforcement learning, plus day-0 support in transformers, SGLang, and vLLM. There’s even a Rust-based harness built for it, which is a nice sign that somebody cared about the plumbing.

Under the hood, Beam is a sparse MoE model: only 8 experts run per token, even though all 256 experts sit in memory. That is how it can behave like a much smaller model at runtime while still carrying the weight of a much larger one. The tradeoff is obvious and very unfashionable: you don’t get to skip the hardware bill just because the active path is smaller. The BF16 checkpoint is about 1 TB, while the NVFP4 version shrinks that to roughly 280 GB. Reflection says NVFP4 is the practical route for serving, with FP8 checkpoints for older GPUs.

The model’s long-context setup is also unusually deliberate. Most layers use a 512-token sliding window, while every sixth layer and the last layer see the full context. That keeps the cache growth down to 36.08 GiB per sequence at full 1M context, instead of 208 GiB if every layer stayed global. Beam also ships with its own tokenizer response template and tool-calling grammar, so parsing its answers and tool calls is meant to be less of a regex hobby and more of a normal software task.

Reflection is dangling $10,000 in Hub credits for the best Beam projects, which is a neat little nudge toward actual usage instead of launch-post applause. And that may be the real point here: Beam is being sold less as a chat toy and more as an open, large-scale system people can deploy, fine-tune, and bend to their own stack.