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

Beam is the first model from Reflection, a new US lab, and this Hugging Face post explains what it takes to run and adapt it: a 501-billion-parameter open-weights model under the permissive Apache 2.0 licence, aimed at coding and agent work for teams that want a frontier-scale model on their own infrastructure. It is a mixture-of-experts model, meaning each token is routed through only a few specialised sub-networks, so only about 23 billion parameters are active at a time even though all 501 billion must be held in memory. (huggingface.co)

Two design choices make the 1-million-token context window practical. Most of Beam's 52 layers only look at a short sliding window of recent tokens, and only a handful attend to the whole input, so the memory cache that grows with context length increases by 36 KiB per token instead of the 208 KiB it would need if every layer kept every token. The weights also come in three precisions: full BF16 at about 1 TB, FP8, and a 4-bit NVFP4 version of about 280 GB for Nvidia's Blackwell GPUs, for which the post recommends four B200s. The post walks through serving with Transformers, vLLM and SGLang, hosted access via Baseten, Fireworks and Together, and fine-tuning examples with LoRA and reinforcement learning.

The headline benchmark figures in the post, such as 81.6% on SWE-bench Verified, were supplied by the Reflection team, and the post does not compare them with other models. An independent evaluation by Artificial Analysis gives a more modest picture: Beam scored 30 on its Intelligence Index, second among US open-weights models behind Muse Spark Open at 39, but only 3.0% on Terminal-Bench 4.0, a newer version of the terminal benchmark than the v2.1 Reflection reports. That evaluation is covered in a separate Intel item.