Reflection releases Beam, a 501-billion-parameter open model

Reflection released Beam’s model weights on Hugging Face under the Apache 2.0 license.[2] Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, and it takes text as input and generates text.[1][3] Reflection says it built the model primarily for coding, reasoning and agentic workloads.[3] It says Beam can be used commercially and for research, including as a base model for fine-tuning.[3] The announcement was published on October 16, 2026.[1]

Reflection says teams can run Beam on their own hardware, infrastructure or cloud, and customize it for their use cases.[1] Developers can download and self-host the model, evaluate and customize it, connect it to agentic systems, or access it through Reflection and its partners.[2] Reflection says Beam is available through an ecosystem of more than 40 partners and integrations.[2] Dell and HPE offer it for on-premises, sovereign and air-gapped deployments.[2] Reflection’s first-party API is free to start through a public beta.[2]

Beam is released in BF16, FP8 and NVFP4 formats.[2] Hugging Face says the model has day-one support in Transformers, SGLang and vLLM, and that Reflection provided recipes for inference, supervised fine-tuning and reinforcement learning.[1] Beam also includes a custom Rust-based harness.[1] Reflection says it pretrained the model on 23.8 trillion tokens from the web and proprietary licensed datasets.[4] Its reinforcement-learning run generated more than 100 million rollouts using 10,500 NVIDIA GB300 GPUs over four weeks, according to Reflection.[4]

Reflection says Beam’s reasoning benchmark scores are comparable to GLM-5.2 while using three to four times less inference compute.[4] It also describes Beam as competitive with larger open models such as GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.[4] Reflection says its advantage over models such as Kimi K3 is inference efficiency rather than raw capability.[4] TechCrunch reported that Reflection’s performance claims had not been independently verified.[5] Reflection says Beam was developed predominantly for English and supports token-efficient reasoning, reaching comparable solution quality to leading open models with about three times fewer inference-time reasoning tokens.[3]