Reflection AI announced Beam on Monday, Oct. 5, as its first open-weight model, pitching it as a Western alternative to models from Chinese labs. The announcement did not make Beam’s weights publicly downloadable, and the company’s performance and computing-efficiency claims have not been independently verified in the material available so far.
For developers weighing models, Reflection has announced a potential option, not yet opened a public release for testing. Beam’s practical value will depend on whether the company publishes the promised weights and documentation, and whether outside evaluations support its comparisons.
What can developers access now?
Reflection said Beam was available to a select group through an early-access sign-up. The company plans to publish the weights later in October under the Apache 2.0 license, along with a technical report, model card and developer materials. It did not give a specific release date. BenchLM’s Oct. 5 profile also reported that the weights were not public at the time.
For developers outside the early-access group, Beam was not yet publicly downloadable. Reflection says the planned materials will help developers run, evaluate and fine-tune the model, but they were not all available at announcement.
Reflection describes Beam as a model for coding, reasoning and agentic tasks. The company reports 501 billion total parameters, with 23 billion active, and says the model was trained on 23.8 trillion tokens. Those figures describe the announced system; they do not establish how it will perform in a developer’s workflow.
What does Reflection mean by a Western answer?
The framing speaks to organizations that want AI systems they can control but may not want, or be able, to use Chinese models. It is a pitch about sourcing and choice, not evidence that Beam is better, safer or more suitable for sovereign deployments. Semafor reported that Chinese technology companies have led the open-source AI field and that some Western firms have used Chinese models for lower-cost routine tasks.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Misha Laskin, Reflection’s co-founder and CEO, told Semafor that organizations seeking sovereign systems do not have many good options today. That is Laskin’s assessment of the market; the reports do not establish how prospective users will choose.
Laskin told Semafor: “It also needs three to four times less computing power to reason through a problem than comparable open models.” That claim could matter to deployment decisions if confirmed, but no independent measurement of it was established.
Do the benchmark results support the comparison?
Reflection’s launch table reports mixed results against named competitors. On DeepSWE v1.1, it lists Beam at 44.4, compared with 44.0 for GLM 5.2, 68.0 for Kimi K3 and 51.0 for Qwen 3.8 Max. On SWE Bench Pro v2-Hard, Reflection reports 77.2 for Beam, below GLM 5.3 at 84.3 and Kimi K3 at 88.2. On AutomationBench public, Beam’s reported score is 37.0, compared with 26.2 for GLM 5.2, 48.2 for GLM 5.3, 46.7 for Kimi K3 and 39.8 for Qwen 3.8 Max. Reflection’s announcement presents these as its own results, not independent replications.
BenchLM labels the scores provider-reported preview results. Its profile said the technical report, model card and evaluation settings were pending, and it did not assign Beam a public rank. Until those details and outside tests are available, developers have limited grounds to judge how closely the comparisons match conditions they care about.
Reflection’s table therefore does not establish that Beam matches or beats Chinese models overall. The announcement also does not establish developer uptake or savings from the claimed efficiency. The next test is whether Reflection publishes the weights under the promised license with documentation that lets others examine its results. For developers, that evidence and performance on their own tasks will matter more than the Western label.