Beam 1.0 501B A23B is a Reflection text model for programming, reasoning and tasks divided into several connected steps. Its sparse MoE architecture has 501.3 billion total parameters and approximately 23.4 billion active parameters per token. The model was developed primarily for English; consider this when choosing the language of source code, technical material and detailed instructions.
The architecture combines local and global attention, and the model’s preparation includes source code, technical documentation, scientific material and educational content. A separate training stage developed long-context capabilities using structured repositories, extended tasks and large documents. During reinforcement learning, the developer combined programming, terminal-work and STEM tasks.
Reflection highlights software engineering, mathematical and scientific reasoning, and sequential problem-solving. These areas lend themselves to substantive text conversations: explaining relationships between program components, investigating a logic error, proposing an algorithm change, planning a hypothesis test or summarizing technical material.
Beam places particular emphasis on efficient use of reasoning tokens. Training rewarded successful solutions that reduced unnecessary tokens. This describes the developer’s approach; you should still specify the desired response length in your request—for example, a brief conclusion, a detailed proof or a patch with an explanation.
The connection has a context window of 1,048,576 tokens, shared by supplied material, instructions, history and the response being generated. The model’s metadata lists output of up to 131,072 tokens. This capacity lets you provide several connected texts: a specification, implementation fragments, verification results and questions about them. Give the material names and ask the model to identify the passage supporting each conclusion, making extensive analyses easier to check.
For a coding task, for example, first ask for a dependency map based on the supplied files, then identify the location of a change and only then prepare the implementation. For a scientific task, specify the initial conditions, notation and required level of rigor. These request structures make it possible to compare a useful result with the source material.
The connection accepts text and returns text through Messages, Responses, Chat Completions and text-based generateContent. Code and recommendations arrive as the model’s response; command execution and access to external data depend on your application’s capabilities.
", "aboutText": "### Beam 1.0 from Reflection\n\nBeam 1.0 501B A23B is a Reflection text model for programming, reasoning and tasks divided into several connected steps. Its sparse MoE architecture has 501.3 billion total parameters and approximately 23.4 billion active parameters per token. The model was developed primarily for English; consider this when choosing the language of source code, technical material and detailed instructions.\n\nThe architecture combines local and global attention, and the model’s preparation includes source code, technical documentation, scientific material and educational content. A separate training stage developed long-context capabilities using structured repositories, extended tasks and large documents. During reinforcement learning, the developer combined programming, terminal-work and STEM tasks.\n\n### The model’s intended tasks\n\nReflection highlights software engineering, mathematical and scientific reasoning, and sequential problem-solving. These areas lend themselves to substantive text conversations: explaining relationships between program components, investigating a logic error, proposing an algorithm change, planning a hypothesis test or summarizing technical material.\n\nBeam places particular emphasis on efficient use of reasoning tokens. Training rewarded successful solutions that reduced unnecessary tokens. This describes the developer’s approach; you should still specify the desired response length in your request—for example, a brief conclusion, a detailed proof or a patch with an explanation.\n\n### A large text context in Ailoot API\n\nThe connection has a context window of 1,048,576 tokens, shared by supplied material, instructions, history and the response being generated. The model’s metadata lists output of up to 131,072 tokens. This capacity lets you provide several connected texts: a specification, implementation fragments, verification results and questions about them. Give the material names and ask the model to identify the passage supporting each conclusion, making extensive analyses easier to check.\n\nFor a coding task, for example, first ask for a dependency map based on the supplied files, then identify the location of a change and only then prepare the implementation. For a scientific task, specify the initial conditions, notation and required level of rigor. These request structures make it possible to compare a useful result with the source material.\n\nThe connection accepts text and returns text through Messages, Responses, Chat Completions and text-based generateContent. Code and recommendations arrive as the model’s response; command execution and access to external data depend on your application’s capabilities.", "locale": "en", "methods": [ { "html": "POST /v1/messagestexttextPOST /v1/responsestexttextPOST /v1/chat/completionstexttextPOST /v1beta/models/reflection/beam-1.0-501b-a23b:generateContenttexttext