Reflection releases Beam, a 501-billion-parameter open model under the Apache 2.0 license

Reflection, a new American AI lab, released Beam, its first model, as open weights on Hugging Face on October 16, 2026. Anyone can download the files and run the model on their own servers or in their own cloud. Beam is aimed at programming and agentic work, meaning tasks where the model plans several steps and calls tools by itself. It accepts only text, handles up to 1 million tokens of context, and was developed primarily for English.

What it takes to run

Beam is a mixture-of-experts model. Its parameters are split across 256 specialized subnetworks per layer, called experts, and a router picks eight of them for each token. Only about 23 billion of the 501 billion parameters are used per token, which keeps generation fast. Memory does not shrink the same way, because every expert has to stay loaded.

The weights come in three precisions. The full BF16 checkpoint is about 1 TB. The NVFP4 build stores most weights in 4 bits and comes down to about 280 GB, but only NVIDIA's Blackwell GPUs run that format natively. For older cards such as the H200, the FP8 build is the one to use. Hugging Face's guide says two B200 GPUs with 180 GB each can just hold the NVFP4 weights, and it recommends four for room to handle long contexts and parallel requests.

Transformers (version 5.20.1 or later), vLLM, and SGLang support the model from day one. If you lack the hardware, Hugging Face points to its Inference Providers service for trying Beam without your own servers. Reflection also ships an open-source harness written in Rust, the program that drives the model through coding tasks.

What the license allows

Apache 2.0 is a permissive open-source license that allows commercial use, modification, and redistribution. Reflection's model card adds that it does not prohibit distillation, training other models on Beam's outputs, fine-tuning, quantization, or merging. The release includes recipes for supervised fine-tuning with TRL and for reinforcement learning with Miles, so teams can adapt the model to their own codebases.

How it scores, by Reflection's own numbers

Reflection reports 81.6% on SWE-bench Verified, a test built from real GitHub issues, and 80.1% on Terminal Bench 2.1, which measures work in a command-line environment. On GPQA Diamond, a set of graduate-level science questions, the company lists 90.6%. These figures are for the BF16 checkpoint, and the model card lists separate results for the FP8 and NVFP4 builds.

The company also says Beam reaches answers of comparable quality to leading open models with about three times fewer reasoning tokens. If that holds, each answer costs less and arrives sooner. All of these numbers come from the publisher, and no independent benchmark had been published by October 16, 2026.