Reflection AI’s table scores Beam at 80.1 on Terminal Bench 2.1, below Kimi K3 at 88.3 and DeepSeek V4.1 Flash at 90.6. That does not show Chinese open-weight models have been displaced: it is a comparison published by Beam’s maker, and AIEvals found no independent Beam results on the benchmark trackers it checked.
Beam should be judged by reproducible performance, full deployment costs and actual availability—not by whether its maker presents it as an answer to Chinese models. Those tests matter to anyone choosing a model, whether their priority is capability, cost, safety or control.
Reflection announced Beam on Monday as its first open-weight model, aimed at coding, reasoning and agentic tasks. The company says it plans to publish the model’s weights under an Apache 2.0 license, along with a technical report, model card and developer materials, later in October. The weights are not public now: Reflection said Beam was undergoing final red-teaming and evaluation, with preview access limited to a waitlist, and gave no specific date for the planned release. (Reflection’s announcement)
A benchmark table is a starting point, not proof
Reflection describes Beam as competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. The company also says Kimi K3 remains ahead on raw capability. Those are Reflection’s characterizations of its own comparisons, not independent rankings.
The table gives specific results, but it does not provide a complete comparison across every model and task. It lists Beam at 77.2 on SWE Bench Pro v2-Hard, for example, compared with Kimi K3 at 88.2 and GLM 5.3 at 84.3; some other entries are marked as unreported. AIEvals says it found no independent Beam results in the Vals AI, Artificial Analysis, DataCurve or Scale AI leaderboards it checked. That finding does not prove no outside testing exists anywhere. It does mean those trackers do not independently confirm Reflection’s reported scores. (AIEvals’ Beam benchmark review)
The distinction matters. A score can help developers decide what to test, but it cannot settle how a model will perform across deployments. Beam’s size—501 billion total parameters, with 23 billion active at a time, according to Reflection—does not by itself establish that it is more useful or cheaper than alternatives.
The same caution applies to the company’s efficiency pitch. Reflection estimates that Beam uses three to four times less inference compute than comparable open models on reasoning tasks. Inference is the computation involved in generating a model’s output. Reflection bases its estimate on active parameter counts and average generated tokens, while excluding prompt prefill, context-dependent attention operations and serving overhead. It is not a measurement of total cost or runtime, and independent tests have not established the claimed savings.
That estimate could matter to organizations paying to run models. But the useful comparison is not an isolated compute figure: it is the cost of serving models under comparable conditions while they complete tasks at comparable quality. Until independent tests establish those figures, the efficiency case remains a claim, not a demonstrated advantage.
The national-security case does not settle performance
There is a serious argument for seeking alternatives to Chinese models. Reflection co-founder and CEO Misha Laskin told Semafor that some businesses and governments want AI systems they can control without relying on Chinese models. Laskin told Semafor, “They don’t really have very good options today.” Semafor also reported that U.S. and U.K. government evaluators found recent Chinese models could help hackers exploit code vulnerabilities. (Semafor’s report on Beam)
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Those concerns can inform procurement. They do not show that Beam is safer, more capable or less costly. The reported security findings concern other models, not Beam; Reflection says it will publish its safety-evaluation results in a future technical report. A model’s national origin is not a substitute for evidence about its behavior and risks.
Nor does the sovereignty rationale establish demand or adoption. Organizations may value control over the systems they use, but whether Beam meets that need depends on its license, deployability, safety and performance once the weights and supporting materials are available. A planned license and a waitlist preview are not yet a broadly usable model.
The test now belongs to Reflection and outside evaluators
Reflection should follow through on its planned release, then publish enough information for independent evaluators and prospective users to reproduce its results. That means disclosing evaluation methods and testing performance, safety and full serving costs—not relying only on selected benchmark scores or an estimate that leaves out important costs.
Outside evaluators should run comparable tasks under disclosed conditions. Organizations considering deployment should check whether Beam delivers the quality they need at a cost and risk they can accept. Procurement officials should apply those standards to models regardless of where they come from.
Beam may become a useful alternative. Its announcement does not establish that it is one. Public weights and supporting materials, independently reproducible evaluations, and cost comparisons that account for the full serving process would give model users grounds to judge.