Reflection: Beam 1.0 501B A23B

Reflection: Beam 1.0 501B A23B is a text model from Reflection. It is an open-weight reasoning model built for coding and long-horizon agentic workflows, and it uses a sparse mixture-of-experts architecture with 501B total parameters and 23B active parameters per token. The provider released it on 2026-10-16, and its knowledge cutoff is 2026-06-30.

The model takes text in and returns text out. Its context window is 1,048,576 tokens, and the maximum output is 943,718 tokens, so it can hold very large inputs and produce very long responses in a single request. Reasoning is part of the model's design: the supported parameters include reasoning, reasoning_effort, and include_reasoning, which let you control whether reasoning is returned and how much effort is spent on it. It also supports tool use through tools and tool_choice, and structured outputs through response_format and structured_outputs.

Those capabilities point to a few practical uses. Long-horizon agentic workflows fit the large context and tool support, since an agent can carry a long history of steps and call tools across many turns. Coding tasks fit the reasoning focus, whether that means working through a large codebase in context or generating and revising code over long outputs. The structured output and tool parameters also suit pipelines that need machine-readable responses or orchestrated calls to external functions. Because the context window is large, tasks such as summarising or querying long documents are viable without chunking, and the high output ceiling helps when a single answer must be extensive.

You call the model through api14's OpenAI-compatible API using the model id "reflection/beam-1.0-501b-a23b". The request parameters it supports are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p. Sampling can be tuned with temperature, top_p, top_k, and min_p, while penalties, seeds, stop sequences, and log probability options cover reproducibility and output shaping.

It is a text-only model, so it does not accept image or audio input, and its output is text only. If your task depends on another modality, a different model would be a better fit. The knowledge cutoff of 2026-06-30 also means anything after that date is outside its training data, so for very recent information you would need to supply it in the prompt or use tools.