Reflection introduces Beam, an open-weight coding and agent model whose efficiency claims remain unverified

What Reflection announced and what is known about availability

Reflection AI introduced Beam as its first open-weight model, designed for coding, reasoning, and agentic tasks that use tools. The company describes it as a mixture-of-experts, or MoE, model with 501 billion total parameters and 23 billion active per token. These specifications come from the company’s announcement; they are not, by themselves, an independent assessment of the model’s capabilities.

In its announcement, Reflection said Beam was undergoing final evaluations and safety testing, and that users could request early access. It also said it planned to release the weights, a technical report, a model card, and developer materials later in October 2026. The evidence provided does not verify the exact release date or confirm that the weights are publicly available. They should therefore not be described as an existing public download.

The announcement says the weights are intended to be released under Apache 2.0. However, the available materials do not include a published copy of the weights or the terms of a license already attached to a distribution. Apache 2.0 should be treated as an announced plan until it can be checked against the repository or the files actually released.

What active parameters mean in an MoE model

An MoE model divides its network into groups of specialized parameters, known as experts, and uses a routing stage to select which experts participate in processing each token. For Beam, the figure of 501 billion describes the model’s total parameter count; the 23 billion active figure describes the portion Reflection says participates in the computation for each generated token. These are different measurements, not interchangeable ways of stating the model’s size.

Activating fewer parameters per token can reduce some of the computation involved in generation compared with an architecture that uses all its parameters at every step. It does not mean the full model takes up only the memory associated with 23 billion parameters. Running it also involves storing the weights, numerical precision, working memory, context length, hardware, and implementation. Without those details and measurements, the active-parameter figure does not establish total operating cost or production speed.

Parameter counts alone are also not enough to judge quality. Results depend on training data and methods, architecture, tool quality, and evaluation conditions. The active-parameter figure helps explain a possible reduction in compute per token, but it needs to be considered alongside comparable results and costs measured on a real system.

What the comparisons show—and where their limits lie

Reflection focuses its capability claims on coding, reasoning, and agentic tasks. In the table published by the company, Beam scores 80.1 on Terminal-Bench v2.1; the figures shown there are 81.0 for GLM 5.2 and 86.6 for Qwen 3.8-Max. On SWE-Bench Verified, the table reports 80.9 for Beam and includes other models’ results where available. These are figures reported by Reflection, not a universal ranking of models.

Results are comparable only when the exact model versions, benchmark and benchmark version, enabled tools, token limits, number of attempts, and scoring method are held constant. The provider’s table also notes an update dated October 8, 2026. A result should therefore be tied to the relevant version and date; figures from different tables or model versions should not be combined casually.

The evidence provided contains no independent reproduction of Beam’s results. The primary announcement documents what Reflection claims, but it does not replace an external evaluation. A strong coding benchmark result also does not establish an advantage across all coding, tool use, or the reliability of agents in deployment. Conclusions should be limited to the tasks and conditions actually tested.

The lower-compute pitch is an estimate, not a serving-cost measurement

Reflection claims Beam achieves scores comparable to GLM 5.2 on advanced reasoning tests while using three to four times less inference compute. The company says it estimates generation operations by multiplying the average number of tokens generated per attempt by roughly twice the number of active parameters. The estimate counts both reasoning tokens and the final answer, and uses third-party evaluation data for other models.

Reflection’s own description of the method excludes prompt prefill, context-dependent attention operations, and serving overhead. The announced ratio is therefore not a comprehensive measurement of inference cost or an operating bill. Nor does it show that a business will pay three to four times less: costs depend, among other things, on infrastructure, utilization, request length, and deployment configuration.

Validating the efficiency claim would require reproducible measurements under the same tasks and conditions: comparable inputs and outputs, identified model versions, documented hardware and numerical precision, context limits, concurrent load, and accounting that includes compute, latency, and operations. Until then, the careful conclusion is that Reflection reports a favorable estimate under a limited method—not that production efficiency has been demonstrated.

Open weights, licensing, and how to follow the announcement

The term “open-weight” describes a way of distributing a model; it does not, by itself, guarantee that every component is open or that use is unrestricted. File availability, the applicable license, documentation, and terms of use need to be checked separately. In this case, Reflection announced that it intended to release the weights under Apache 2.0, but the evidence provided does not confirm that a licensed version has been published.

It is also useful to distinguish a hosted API from access to model weights. An API lets users send requests to a service operated by the provider; downloading weights may allow a model to run on an organization’s own infrastructure, subject to the relevant technical and legal requirements. The announcement describes early access by registration and a later plan for weights and tools, but it does not establish that all those options are already open to the public.

The priority checks for an update are whether the files have actually been released; the exact license text accompanying them; a technical report and model card identifying versions and configurations; and independent evaluations of coding, reasoning, and tool use. Until that evidence exists, the capabilities and compute savings should be attributed to Reflection, and availability should be described as announced rather than completed.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.