Reflection AI first announced Beam on October 5 and promised the weights later in the month. On October 16 it delivered. Beam 1.0 is now on Hugging Face under Apache 2.0, which means you can use it, change it and build commercial products on it without a separate license.
According to the model card, Beam is a mixture-of-experts model with 501.3B total parameters and 23.4B active, spread across 256 experts and 52 layers. It handles text only, was built mostly for English, has a 1M-token context window, and its training data runs to June 30, 2026. Reflection released BF16, FP8 and NVFP4 versions, along with the tokenizer, a reference implementation and a DFlash2 draft model for faster decoding. The company says it trained Beam for coding and agent work, and claims it reaches similar quality to leading open models while using about three times fewer reasoning tokens.
The part that stands out is Mirror CLI. Reflection describes it as the open-source harness it built and trained Beam on, so you get the model and the agent loop it was tuned for at the same time. The company says Beam is available through more than 40 partners.
The first third-party rankings are more modest. On Arena, Beam ranks 12th among open models in Agent Arena, with a margin of error of ±3.84%, and 22nd among open models in Code Arena WebDev with 1,438 points. On both boards it is the highest-ranked US open model. Artificial Analysis gives it 30 on its Intelligence Index, second among US open-weights models behind Meta's Muse Spark Open at 39, and calls Terminal-Bench 4.0 a weak spot, where Beam scored 3.0%. Reflection says a technical report is coming.