Reflection's Beam model weights became available on Hugging Face on October 16, 2026. Hugging Face lists the release under Apache 2.0, while Baseten announced same-day access through its Model APIs and Dedicated Inference. This is the weights-and-serving release, distinct from Reflection's October 5 introduction.
The published model has 501 billion total parameters and activates 23 billion parameters per token. It accepts text and produces text, mainly in English, with a context window of up to one million tokens. The sources describe coding, reasoning and agentic workloads; they do not establish current image or audio input support.
Hugging Face lists BF16, FP8 and NVFP4 checkpoints, along with support in Transformers, vLLM and SGLang. The release also includes a Rust-based agent harness. Available integration recipes do not mean every existing application or inference server is ready without configuration.
The smaller active-parameter count does not make this a small-memory model. Hugging Face says all experts remain in memory: the BF16 checkpoint is about one terabyte, and the NVFP4 checkpoint is about 280 gigabytes. Its NVFP4 deployment guidance targets NVIDIA Blackwell GPUs, so hardware and precision support need checking before self-hosting.
Baseten describes adjustable reasoning effort and training intended to reduce unnecessary tokens. Published performance tables are supplied by Reflection and vary by checkpoint precision. They are not an independent guarantee of task accuracy, lower total cost or safe tool use in a buyer's workflow.
Teams comparing hosted access with self-hosting should check current provider prices, capacity, data handling and operational controls before committing. The intake draft's specific token-price and benchmark figures are omitted here because they were not confirmed by the primary pages used for this brief. Open weights provide a deployment option, not automatic privacy, security or production readiness.