Reflection Beam: Coding and Agent Model in Selective Preview
Reflection Beam is a text-only model for coding, reasoning and agent workflows. Access is selective; model-level context claims and current beta API limits differ.
Checked October 6, 2026 · Documentation review, no hands-on product test
A coding model, with access still limited
Reflection reports a sparse mixture-of-experts architecture with 501B total parameters and 23B active parameters, plus 1M effective context. The model is text-only; using OCR or another tool to convert material to text does not make its native input multimodal. These are developer-reported specifications, not our measurements. Official launch report.
Beam is intended for code and multi-step agent work. It is distinct from a dedicated decision model returning a bounded classification or score. For that narrower job, use the Decision Models selection guide.
Check the limit of the actual API
The current beta model documentation lists Beam-501B-A23B with a 256K total context window and 128K maximum output. Input and output share the window; beta limits may change. The launch report's 1M effective context is not a promise that the present endpoint accepts a 1M request. Current API model limits.
The API documents reasoning, tool calling and structured outputs. Its OpenAI-compatible route covers Chat Completions and Models, not Responses, Images or Audio. Check the exact endpoint and client settings rather than assuming every SDK default works. Compatibility scope.
Preview access and planned weights
The developer introduction describes beta access through a waitlist; a public playground or documentation page does not establish account eligibility. Access documentation · Platform.
Reflection plans to publish weights, a technical report, model card and developer artifacts during October 2026, with an Apache-2.0 weight release. As of this check, that remains a release plan; downloadable Beam weights were not verified through the official links. Release statement · Developer-linked organization.
Public API prices, self-hosting hardware requirements and our account access remain unconfirmed. Active parameter count describes work per token, not all memory needed to hold weights, context and serving overhead.
A useful evaluation brief after access is granted
Use a permissioned repository task with a known acceptance check: explain a call path, propose a small patch, then inspect the diff and run the relevant checks. Record the model ID and effective API limit. Keep tool permissions and deployment approval outside model output.
This is a proposed evaluation, not a Beam result. Avoid choosing it on launch benchmark comparisons alone or assuming preview access is suitable for a production dependency. Browse Coding & Development for other development workflows.
Evidence and next step
Sources are linked beside the claims they support. Workflow suggestions are editorial examples, not completed tests; no account, installation, paid generation or benchmark was used for this profile.
Start with the official product or project and confirm the exact access, version and terms needed for your task. How we review.