Reflection announced downloadable Beam weights under Apache 2.0 on 16 October 2026. The model targets coding, reasoning and extended tool-using tasks. The company also presents Mirror CLI as a separate environment for agent work. This release combines a model with ways to use it; availability of one component does not establish the quality of the entire system.
What the model contains
The published model card describes a mixture-of-experts architecture with 501.3 billion total and 23.4 billion active parameters, a context window of up to one million tokens, and text input and output. Development focused predominantly on English. These are manufacturer-supplied characteristics, rather than SIExpo measurements.
The public repository describes the instruction-tuned release in BF16, FP8 and NVFP4, alongside a modified Harmony conversation format. Numerical precision and the input template belong to the configuration. A result reported for one variant should not automatically be assigned to another.
Our technical interpretation is that active parameters describe the portion used in a particular computation, rather than the complete deployment footprint. A team still needs space for the model, serving software and context memory. A long conversation and several concurrent users place different demands on a system from one short prompt. The 23.4 billion figure alone therefore cannot serve as a hardware recommendation.
The agent is another layer
Mirror's documentation describes workspaces, a visible activity stream, conversation management and tool approvals. Users can steer work while it runs. This provides a concrete interface between an intention and an action; it does not establish that every edit is correct or that every granted permission is appropriate.
The guide explicitly says approval mode is not a sandbox: commands and extensions run with the user's account permissions. Rewinding a conversation also does not undo files already edited. These are concrete boundaries an operator needs to consider when choosing an execution environment.
Consider a hypothetical task to repair an import in a small project. The model proposes a change, the environment writes it, and a separate check establishes whether the application starts and the original problem is resolved. Those three outcomes need to remain distinguishable. A detailed agent report cannot replace inspection of the actual file.
Extended work also raises the question of unfinished steps. If a conversation stops after an edit but before a test, resuming should preserve that distinction. In a proposed pilot, we would compare the agent's account with the changes and the results actually available. This illustrates an evaluation method; we have not run Mirror or Beam against such a project.
What operating your own endpoint means
Downloaded weights provide a choice about where the model runs. Self-hosting adds responsibility for versions, capacity, monitoring and recovery. Control of a machine does not automatically decide who can read conversations, start tools or receive logged data. Those decisions belong to the application around the model, including its operational policies and access controls.
For a Bulgarian workflow, we would assemble representative cases containing the relevant terminology, code and expected outcome. The predominantly English development focus is a reason to evaluate, rather than a predetermined verdict. A useful measure is successful task completion, including human corrections, rather than the apparent confidence of the initial answer.
Reflection supplies the published benchmark evaluations; we do not present them as an independent ranking. The distinction between a demonstration and a useful product appears in repeatability: the same instructions, versions, permissions and criteria should make comparisons meaningful. Recording only a final screenshot would lose much of that context, even when the visible result looks convincing.
An open release makes that kind of internal evaluation possible, but does not perform it for the adopting team. Beam's practical change today is the availability of a model that organizations can inspect operationally and place within their chosen environment. Whether that choice improves a particular workflow remains a question for evidence from that workflow, with the model and its harness evaluated together.
Editorially checked: 2026-10-16T16:26:59.575073+00:00 (UTC). Cover: conceptual AI illustration.
