Scorecard · International
Reflection actually releases Beam’s weights, moving beyond its preview
View the linked event assessment →
On October 16, Reflection released its first model, Beam, for public download under Apache 2.0. The release includes versions stored at different numerical precision, documentation and integrations for several inference systems. Developers can possess the model, modify it and run it on infrastructure they choose, rather than obtain every response through Reflection’s service. The public repository and model card ↗ identify the checkpoint and licence; the compressed version ↗ is also listed.
Beam contains approximately 501 billion parameters, of which about 23 billion participate in processing each token. The smaller active number reduces computation; it does not make the entire model small enough for an ordinary laptop. Storage, memory, engineering and hosting costs remain substantial. Provider claims about performance and the breadth of partner availability have not been independently tested here. Different serving software is a choice of implementation, not proof that every listed partner has an operating endpoint.
The current landscape’s D1006-07a correctly records an October 5 preview and a promise of later weights. This release supplies the missing ownership/deployment option. It is a separate model from the already scored October 15 Meta release; it is not another award for that event.
Approved disposition: Promote D1006-07a to a Specific Event and Dispersion +1, dated October 16; Leaning dispersion / Opens; lanes 4/5, with 7 supporting. Recognize legal and practical possession of an alternative model, while reserving a stronger judgment for independently demonstrated use. TR-015/TR-018 context only, excluded from trend scoring counts. No new Policy Tool for a model release. Review independent deployment and dependencies October 30.
Follow-up: October 30: independent deployment, hosting/hardware costs and dependencies.
November 3, 2026 — evidence update
An October 27 VulcanBench report tested Reflection Beam on 23 software-reconstruction tasks at five reasoning settings—115 runs through Reflection's API and Mirror CLI. It reports no fully solved tasks at the first three settings, four at extra-high and six at maximum effort. More reasoning generally improved functional performance, while increasing time and token use. Two other models judged code quality; hidden tests measured behavior. These are this evaluator's results, not a representative survey of production work.
Crucially, this was not a self-hosted deployment test. It supplies an external performance check but does not resolve the hardware floor, independent serving cost or practical autonomy of a smaller institution. The current D1006-07a already has a +1 for the October 16 weight release; this evaluation neither repeats that award nor establishes that possession alone delivers usable capability.
Supporting sources
Source snapshot: 2026-11-03. Individual event and verification dates remain as recorded above. Inclusion alone does not assign a score.
