Software companies and engineering teams can now download Reflection’s Beam coding model to customize it and run it on computing infrastructure they control. Hugging Face announced the public-weight release on October 16, 2026.
Weights are the files containing a model’s learned parameters. Access to them lets a team operate and adapt the model, rather than only send requests to a provider. Reflection’s October 5 preview announcement had promised the downloads for later that month; the new development is their public availability.
Beam’s Apache 2.0 licence permits commercial use, modification and redistribution, subject to conditions including preserving applicable notices and identifying changes. Reflection’s model card expressly allows fine-tuning—adapting the model through further training. It describes Beam as text-only and developed predominantly for English, with a focus on coding and tasks involving software tools.
Running Beam is a server project
Reflection offers three numerical formats for storing the model: BF16, FP8 and the more compressed NVFP4. Its download guide lists the NVFP4 files at 273 GiB. That is the download’s storage size, not the total memory needed to run the assistant.
The architecture helps explain the distinction. Beam has about 501 billion parameters, but uses about 23 billion to process each token, a small unit of text. Hugging Face explains that the unselected parts still occupy memory in the documented deployment—using less of the model at each step does not make the complete model small enough for an ordinary workstation.
Reflection’s reference serving setup uses four NVIDIA B300 GPUs in one server. It is explicitly an example for development and evaluation, not a universal hardware minimum or a production-capacity guarantee. Memory must also accommodate the requests being processed; the weight-file size alone cannot determine how many people the service can support. Hugging Face’s hardware guidance makes that extra capacity requirement explicit.
Decide what control is worth
Running the model yourself is not the only route. The Hugging Face release post also lists hosted access through Baseten, Fireworks and Together. That offers a way to evaluate Beam without first operating a model server; account access and provider terms still need checking.
A coding application installed on your computer is not necessarily using a local model. Reflection’s Mirror coding agent defaults to Reflection’s provider, while its configuration documentation also supports a self-hosted endpoint. The practical implication: a team seeking to keep model processing on its own infrastructure must check where the application sends requests, not just where it is installed.
Nor do Reflection’s efficiency claims establish a lower hosting bill. The company says its original compute comparison is an estimate that excludes parts of request processing and serving overhead, not a measurement of inference cost.
Our assessment: Beam warrants a self-hosting evaluation when adapting the model or controlling its deployment solves a specific business need. Before committing, obtain a capacity quote for the expected request lengths and simultaneous users, then compare costs against coding tasks that meet your team’s acceptance criteria. The hosting-cost framework in our Step 5 guide explains that comparison; its product-specific prices and licence do not apply to Beam.