Nvidia in Talks to Acquire Reflection AI, a Startup It Already Backed: FT
The Financial Times reported on October 10 that early talks over an acquisition or further investment are under way. Reflection AI has released Beam, its first open model, with trial access limited to a select group of users.
Nvidia is in talks to acquire or make an additional investment in the US AI startup Reflection AI, the Financial Times reported on October 10. The discussions are at an early stage, and hiring and technology licensing arrangements are also on the table, according to the FT. Reuters said it could not independently confirm the report.
Already an investor, now a possible buyer
The two companies are already tied together through investment and model development. In an official post in 2025, Reflection AI named Nvidia as one of its investors. Nvidia, for its part, announced in March 2026 that Reflection AI had joined the “Nemotron Coalition,” a group jointly developing open models. The latest report concerns whether that partnership will expand into a larger investment or an outright acquisition.
The shape of any deal has yet to be settled. Alongside an acquisition of the company, the FT report mentions additional investment, hiring, and technology licensing. Reuters said neither company immediately responded to requests for comment at the time of the report.
What kind of model is Beam?
Reflection AI introduced Beam on October 5, five days before the reports of acquisition talks surfaced. Billed by the company as its first open model, it targets coding, problem solving, and the use of external tools. The company said it plans to release Beam’s model weights — the values produced by training — sometime in October. Beam is a text-only model.
Beam is built to activate only the parts it needs. The model totals 501 billion parameters, but only 23 billion are used to generate any single response. The company said the additional training that shapes how the model answers used 10,500 Nvidia GB300 GPUs over four weeks. Nvidia is both an investor and the supplier of the hardware used to train Beam.
In benchmarks run by the company itself, Beam scored 80.9 on a software problem-solving test (SWE-bench Verified) and 80.1 on a test of computer command handling (Terminal-Bench 2.1). The company said Beam delivers problem-solving ability comparable to GLM 5.2 while using less compute. It added that the compute comparison is an estimate rather than a measurement of actual serving costs.
Can you try Beam right now?
As of October 10, the early version of Beam is available only to a selected group of users. The company is taking sign-ups for a waitlist. The weights, a technical report, model documentation, and tooling for running it are all slated for release during October. The company plans to release it under the Apache 2.0 open-source license.
There is a Korean connection as well. Reflection AI has posted plans on its website to build AI cloud infrastructure for Korean companies and institutions. For readers in Korea, the question alongside the fate of the acquisition talks is when Beam will be usable and on what terms. The next things to watch: whether the promised model release lands on schedule, how well it performs on Korean-language tasks, and what it actually costs to use.