Reflection Releases Beam, a 501B Open Model for Local AI

Created
by Ben Voss · Acceledesk AI news writer
Reflection Releases Beam, a 501B Open Model for Local AI

Reflection has released Beam, a 501-billion-parameter open model for coding, reasoning, and AI agents. The model weights are available on Hugging Face under the Apache 2.0 license. Reflection says Beam is also available through its service and a network of launch partners.

501B Parameters, 23B Active

Beam uses a sparse mixture-of-experts design. It contains about 501 billion total parameters, but activates about 23 billion for each token. This allows each response to use only part of the full network, according to Reflection’s availability announcement and the model card.

Beam is a text-only model with a context window of up to 1 million tokens. Its released versions include BF16, FP8, and NVFP4 formats for different hardware platforms.

Apache License Opens Deployment

The Apache 2.0 license allows commercial use, modification, and redistribution subject to its terms. That gives organizations more control over deployment and customization than a model available only through a closed hosted service.

Reflection says Beam launched with more than 40 partners and integrations, including cloud providers, inference platforms, and open-source projects. The company says developers can download the weights, self-host the model, evaluate it, and connect it to agent systems.

Coding Agents Get New Option

Beam targets software engineering, reasoning, and long-running agent tasks. Its model card lists coding, browsing, tool use, and Model Context Protocol support. Reflection also says Beam was trained for efficient reasoning, using fewer inference-time tokens than comparable open models.

The model card reports an 81.6% score on SWE-bench Verified for the BF16 version, along with results on several reasoning, coding, and tool-use evaluations. These results are reported by Reflection and displayed in the model card.

Hardware Limits Remain

Beam’s sparse design does not make it a small model. Hugging Face says all of its experts remain in memory during operation, and the BF16 checkpoint is about 1 terabyte. The full model is therefore aimed mainly at AI companies, research groups, and developers with access to substantial computing resources.

Beam gives developers another alternative to closed AI services, while its open license and large context window support more flexible deployment and customization.

References