Nvidia is in early talks to buy Reflection AI or increase its investment in the startup, according to a Financial Times report on October 10. Reflection introduced Beam five days earlier.
Reuters, citing the Financial Times, reported that one option would involve Nvidia hiring employees and licensing technology. An agreement could still fall apart. Neither company had announced a transaction as of October 11.
Reflection’s chief executive, Misha Laskin, has made control over AI a central part of his pitch. “Closed models are the equivalent in real estate to renting an apartment,” he told Alex Heath in an interview published on October 5.
Developers who download a model can keep using their copy when a supplier changes its product or business plans, subject to its license. Companies using a hosted service depend on continued access to that service.
What developers would receive
In its October 5 announcement, Reflection promised to publish Beam’s weights under the Apache 2.0 license later this month. Weights are numerical settings learned during training. Alongside those files, the company plans to release documentation and software for running, evaluating and adapting Beam. Beam has 501 billion parameters, with 23 billion active for each token it processes, according to Reflection.
Apache’s copyright grant is perpetual and irrevocable, subject to compliance with its terms. A recipient can copy, modify and distribute covered material without paying a royalty. Selling the company that issued such a license would not, by itself, cancel those permissions for copies already released.
Those terms require users who redistribute material to preserve specified notices, include the license and identify modified files. They do not require a developer to publish future models, maintain a hosted service or provide support. Beam’s promised license would protect use of a released version; obtaining its successor would depend on a separate release.
Reflection says it trained Beam on 23.8 trillion tokens, small units of text, drawn from web content and licensed datasets. Downloading the resulting weights would give a company a model it could run and adapt, while reproducing that training would require access to the underlying material and processing methods.
What the training disclosure reveals
Reflection’s October training-data summary describes roughly 20 trillion unique tokens across its foundation-model datasets. That disclosure covers the company’s data collection, while the 23.8 trillion figure describes Beam’s pretraining. Reflection says it repeated some code and technical content during that training; its documents do not provide a reconciliation of the two totals.
The summary lists internet material, public repositories, purchased or licensed collections, human annotations and data generated inside Reflection. It says collection began in September 2025 and continues. It also acknowledges that datasets may include copyrighted works and incidental personal information, despite filtering intended to reduce sensitive information.
Reflection has described broad sources and preparation methods, but these documents do not identify each licensed collection or its supplier. Engineers trying to understand where a model learned a particular subject would need more detail than these categories provide.
The Open Source Initiative’s definition of open-source AI calls for information detailed enough for a skilled person to build a substantially equivalent system. Its requirements include where training data came from, how developers selected and labeled it, and which public or commercial collections others can obtain. The definition allows descriptions of data that cannot itself be shared.
Where customer conversations would go
A company running downloaded weights on its own infrastructure can choose where prompts and responses are processed. Developers calling Reflection’s hosted model send those exchanges to Reflection, where its published API policy permits retention for safety and abuse monitoring.
Ordinary prompts and responses can be kept for up to 30 days. Flagged content can remain for 12 months, or longer if legally required. Reflection says it does not train models on customer content by default, although it may use safety-related metadata and telemetry for training and evaluation. Users can opt in to training on their content.
Deleting an account leaves the retention period for flagged content in place. Applications built by other developers may also keep their own copies of a conversation under separate policies. Reflection’s deletion process does not remove those copies.
Reflection’s privacy policy, last updated August 28, permits transfers of information in connection with negotiations or completion of a corporate transaction. It excludes content processed for enterprise customers under their separate agreements. A business using that offering would therefore have to consult its contract to determine how a sale could affect its data.
Reflection says it will release Beam’s weights and technical report in October, along with safety results and evaluations developed internally.
