Reflection Beam
501B-parameter, 23B-active open-weight reasoning and coding model; weights and technical report released October 16, 2026 under Apache 2.0.
Score calculation
(6 + 7 + 9) ÷ 3 = 7.3
3 of six dimensions rated. Unrated dimensions and BDS status are excluded.
Full scoring method ↗Setup details
Training and release facts are Beam-specific and unusually detailed. Privacy after self-hosting is controlled by deployment operator; Reflection API terms remain separate. Open weights increase deployment control but also leave downstream uses less observable.
Reflection announced Beam October 5 and released weights, model card, and safety report on October 16, 2026. Checked 2026-10-23.
Energy
↑Reflection reports 6,144 GB300 GPUs for under four weeks of pretraining and 10.5K GB300 GPUs for a four-week RL run, but no kWh, water, or emissions totals. Compute disclosures are unusually detailed but not environmental accounting.
Reasoning & sources
Reflection reports accelerator counts and durations for pretraining and RL, which make compute scale more visible than most releases but still omit facility energy, water, and emissions. Treat these as useful inputs for a future estimate, not as a verified low- or high-carbon footprint.
No measured training energy, water, or carbon total; no inference energy per task.
Training rights
↑Reflection says Beam used 23.8T tokens from web, public sources, and proprietary licensed datasets, and describes filtering/curation. It says almost all publicly accessible and unrestrictively licensed code/docs were included; the page does not list all sources or provide an external rights audit.
Reasoning & sources
Reflection describes a 23.8-trillion-token mixture, public and proprietary licensed sources, filtering, and curation, with further detail in the model card and safety report. This supports a comparatively specific provenance account, but independent confirmation of licenses and coverage remains unavailable.
Reflection discloses public/web and licensed data categories, but no itemized corpus or independent rights audit.
No itemized training corpus, full license manifest, or independent rights audit.
Privacy
↑Beam can be self-hosted under open weights, allowing operators to keep prompts locally. Reflection also offers a hosted API; API privacy and retention need to be assessed separately. No single policy covers all partner endpoints.
Reasoning & sources
Beam’s weights can be self-hosted, while Reflection also offers an API; those routes have different data controllers and retention risks. Use the local option only with operator controls, and review the API terms separately before sending sensitive prompts.
Public weights enable local deployment and operator control over prompts, but that is a deployment option rather than a guaranteed privacy safeguard; hosted services and downstream configurations require separate review.
Need current first-party API terms and partner-specific logging/retention controls.
Openness
↑Reflection released model weights under Apache 2.0 along with model card, safety technical report, serving formats, and deployment/customization artifacts. This is strong weight and technical openness; source-level training data remains only partially disclosed.
Reasoning & sources
Reflection released Apache 2.0 weights plus a model card, safety report, and serving artifacts, supporting substantial reuse and inspection. The training data are described rather than fully reproducible from a public source-by-source corpus, so openness is strong but not complete.
Apache 2.0 weights, model card, safety report, and multiple deployment artifacts are released.
Full training data is not published, so the full training pipeline is not reproducible from public artifacts alone.
Worker treatment
↑Reflection describes human-in-the-loop curation, environments, grading, and expert review but does not disclose data worker count, compensation, vendor chain, or protections.
Reasoning & sources
Reflection describes human curation, grading, and expert review but reports no worker counts, compensation, vendors, or protections. The evidence confirms human work but does not describe worker conditions; request details about the labor chain.
Human involvement is described; labor conditions remain undisclosed.
Government & military
↑Reflection says Beam is available to governments and supports sovereign deployments, but the cited launch page does not establish a military contract or military deployment. No model-specific contract was verified in this pass.
Reasoning & sources
Reflection says Beam can serve government and sovereign customers, but the cited announcement does not establish military procurement or deployment. Treat public-sector availability as a possible use context, not a defense contract; no military deployment is established.
No exact defense procurement or end-use evidence found.
BDS status
↑We did not find Reflection AI among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.
Reasoning & sources
Reflection AI is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well.
This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.