Hugging Face offers 501B Beam inference and one-click deployment

Top News

Beam is Reflection's Apache 2.0, 501B MoE for teams that want to host and customize their model. You can access it through Hugging Face Inference Providers or deploy it through Inference Endpoints.

Here's what changed:

• Hosted inference runs through Hugging Face Inference Providers via Baseten, Fireworks, and Together.
• One-click Inference Endpoints deployments support FP8 or NVFP4, with CLI deployment available too.
• Transformers, SGLang, and vLLM support Beam, with supervised fine-tuning and reinforcement-learning recipes available.
• Each token activates 23B parameters, and the model supports tool calling and five reasoning-effort settings.

One catch: all 501B parameters must stay in memory. Plan on four B200s for long contexts and concurrency. Two B200s hold NVFP4 weights but leave little room for activations and KV cache. H200s cannot run NVFP4 natively.

Try it: install with pip install -U "transformers>=5.20.1", then load reflection/Beam-1.0-501B-A23B with the Transformers text-generation pipeline. For hosted inference, use reflection/Beam-1.0-501B-A23B:auto.

Sources: Hugging Face