Reflection, a new US AI lab, has released its first model, and it is a large one given away on generous terms. Beam is a mixture-of-experts model with 501B parameters, of which 23B are active for each token, a 1M-token context window and an Apache 2.0 licence that allows commercial use and modification. It is built for coding and agent work, and aimed at teams that want a frontier-scale model on their own hardware rather than behind someone else's API.
Hugging Face's write-up gives an unusually detailed look at the design. The model has 52 layers; most use sliding-window attention over the previous 512 tokens, while nine attend to the whole context, which keeps the memory needed for long contexts to about a sixth of a naive design. Each of 49 layers holds 256 experts, eight of which run per token, picked by a router modelled on DeepSeek-V3's. Reasoning can carry across turns, there are five effort levels from low to max, and tool calls use a dedicated grammar. Beam has day-one support in Transformers, vLLM and SGLang, ships with Reflection's own open-source harness written in Rust, and comes with fine-tuning and reinforcement learning recipes.
The hardware is the practical barrier. The BF16 checkpoint is about 1 TB. An NVFP4 version cuts that to about 280 GB, which Hugging Face says fits on two Nvidia B200 GPUs with little headroom; it recommends planning for four. NVFP4 runs natively only on Blackwell chips, so older H200 setups need the FP8 checkpoints, and the vLLM example serves the full model across 16 H200 GPUs. For most teams, that means renting capacity or using hosted providers, which Hugging Face offers.
A permissive licence at this scale is the real news. Many large open-weight models come with restrictions on commercial use or user numbers, and Apache 2.0 removes that question for companies that need to self-host for privacy or control. It joins Meta's Muse Spark Open, released the same week, as evidence that US labs are competing again at the top of open weights. We review Beam alongside the serving tools it supports, vLLM and SGLang.
What is missing is independent proof that it codes as well as frontier models. The benchmark results in the post are supplied by Reflection, and as a first model from a new lab it has no track record. Reflection is offering $10,000 in Hub credits for the best projects built on it, which should bring early community testing. Until independent coding benchmarks appear, treat Beam as a promising self-hosting option, not a confirmed alternative to the leading closed models.