Reflection releases Beam: open weights, coding agents and what to test first

What is available now?

Reflection’s release includes downloadable weights, a model card and safety documentation. The company also lists a first-party API public beta and partner access. Its launch describes more than 40 partners and integrations spanning inference, customization and agent tooling.

The important date distinction: Reflection introduced Beam on October 5; the October 16 announcement marks availability of the weights and wider ecosystem. This article covers that release milestone.

BF16, FP8 and NVFP4 weight formats give deployment teams several precision options. The launch also names inference integrations including Transformers, vLLM and SGLang. Check the actual endpoint or deployment guide before selecting a provider.

Public record: Reflection — Beam is now available, October 16, 2026 · Reflection — Introducing Beam, October 5, 2026

Understand the model before choosing hardware

Reflection’s model card specifies 501.3 billion total parameters, with 23.4 billion active, and a one-million-token context window. It accepts text and returns text. The card says it was developed predominantly for English and gives a June 30, 2026 knowledge cutoff.

Active parameters describe the subset used during computation; they do not mean the full model has the storage footprint of a 23-billion-parameter model. Our assessment: hosted access is the simpler starting point for teams without an established model-serving operation.

Reflection positions Beam for coding, reasoning and multi-step tool use. Those capabilities still require testing on your work. A context-window specification alone does not show how accurately the model will find relevant details in your repository.

Public record: Reflection — Beam model card

Mirror CLI turns the model into a working coding agent

Reflection’s developer documentation describes Mirror CLI as its terminal coding agent. It can inspect code, edit files and run commands, with Beam as the default model. Mirror is in early beta, so commands and behavior may change.

The documentation describes resumable conversations, independent workspaces and connections that extend available tools. It also explains manual and automatic approval modes. Approval mode is not a sandbox: commands and extensions run with the user’s account permissions.

For a first evaluation, use a disposable copy of a small repository with non-sensitive data. Keep permissions narrow enough that an incorrect command remains easy to recover from. This is Promptwire’s proposed trial setup, not a report of hands-on testing.

Public record: Reflection Developer Docs — Mirror CLI overview

Measure accepted work, not the launch demo

Choose three tasks from your own backlog: a reproducible bug, a small feature with written acceptance criteria, and a refactor whose behavior must stay consistent. Give Beam and your current tool the same starting code and instructions.

Record whether the patch passes relevant checks, how much reviewer time it needs, elapsed time and total cost. Include at least one incomplete requirement to see whether the agent asks for missing information or makes an unsupported assumption.

Do not turn a company benchmark into a guarantee about your project. Promptwire has not independently tested Beam. The useful question is whether its accepted changes justify the review effort in your environment.

What to confirm before a larger rollout

Verify current endpoint pricing, access conditions, data handling and operational limits directly with the provider you choose. This article does not quote an unverified token price.

For self-hosting, confirm the selected precision format, infrastructure requirements and ongoing serving costs. Downloadable weights give you another deployment choice; running that deployment remains a separate engineering responsibility.