New Reflection Beam

Beam is an open source weighted model released by Reflection, an American AI company, with a parameter scale of 501 billion. It is positioned as an "efficient agent model" and focuses on tasks such as programming and tool invocation that require multi-step execution.

The weights can be directly downloaded on Hugging Face, and you can fine-tune and deploy them yourself. If you don’t want to build it yourself, you can also apply for key calling on Reflection’s API platform. The company also provides open source deployment containers, customized solutions and agent frameworks, supporting deployment in private clouds, local computer rooms and even offline environments. It is obviously aimed at enterprise customers with data requirements.

Running a 500B level model locally requires a multi-card server, and individual developers can basically only use the API. The actual ability depends on third-party evaluation. You can first use the API to compare your commonly used models.