JOURNAL / 2026.10.17
Reflection releases Beam weights and tools to verify its deployment
Beam moves from selective beta access to Apache 2.0 weights, with inference engines and public tests. The release enables investigation of how much behavior depends on the model and how much on its integration.
On October 6, we wrote about Beam with one central reservation: its weights were still a promise. Reflection released them on October 16, alongside documentation and a safety report. The change is substantial: third parties can preserve, modify and host the model rather than rely exclusively on access granted to a service. The first-party API is still described as a public beta; downloading weights and obtaining service access are separate routes.
The model card describes a text model, developed predominantly for English, with 501 billion total parameters, roughly 23 billion active per token and a million-token context. BF16, FP8 and NVFP4 versions are available. The card publishes evaluations by precision, but still lists the general technical report as forthcoming. The available safety report should not be confused with that architecture and training document.
The weights carry the Apache 2.0 license. It permits redistribution and modification subject to its license and notice requirements, without warranty. For a team maintaining its own adaptation, this offers a clearer basis for continuity than informal permission to access a service.
Hosting requires a specific integration. The recipe maintained by vLLM uses a dedicated image because support has not yet shipped in a published release. It also configures parsers for reasoning and tool calls. This is a useful reminder that downloading the checkpoint is only part of the job: interpreting where an answer ends or an action begins is part of system correctness. Declared API compatibility does not settle that work on its own.
One resource strikes me as particularly valuable here: Beam Verifier. It separates functional tests—chat, tools, structured output, multiple turns and streaming—from capability evaluations with predefined settings. Its documentation identifies agent requests whose tool arguments often break unless the server constrains them to the corresponding schema. This lets developers investigate integration failures before attributing them to weak reasoning. Passing those tests does not certify an entire application either.
The release also includes DFlash2, an auxiliary model that proposes token blocks for Beam to verify. It does not work as an independent assistant. Reflection explains that rejection sampling preserves Beam’s output distribution. This acceleration route changes execution; it does not make the model more intelligent. Published measurements use NVFP4 on B300 hardware and describe generation, not the total time needed to solve a task with tools.
There is an inspectable agent, too. The Mirror documentation describes an early beta and explains its manual approval mode. It specifies that this mode is not a sandbox: approved commands run with the account’s permissions. When evaluating a local agent, that distinction matters as much as its ability to write code. My approach would be to start in a reversible project and limit permissions outside the model.
The safety report rates Beam as “Limited” across the framework’s four systemic-risk categories. This is Reflection’s assessment of marginal risk relative to capabilities already available, rather than a declaration of harmlessness. The document acknowledges that its loss-of-control audits measure propensities in constructed scenarios and that the autonomy evaluation with Andon Labs is ongoing. Publishing methods and limits improves the discussion; it does not settle behavior in another environment.
My assessment has changed since the beta: there is now enough material for the discussion to move beyond demonstrations selected by Reflection. The most interesting next step would be public reproductions that preserve the checkpoint, engine, output constraints and tools. If two deployments differ, locating the difference would be a useful result even without topping a ranking. Openness gains value when it lets people explain and fix a failure, as well as produce a good answer.
Primary sources consulted on October 17, 2026. We have not run Beam, Mirror or its verifiers.