Nvidia is in talks to acquire or deepen its investment in Reflection AI, a U.S. developer of open-weight AI models, the Financial Times reported Saturday. The talks are at an early stage, according to people with direct knowledge of the matter cited by the FT.

One option under discussion is an acqui-hire, in which Nvidia would hire Reflection's staff and license its technology. That structure would avoid the drawn-out regulatory review that comes with a full acquisition, the people said. Nvidia could also take a larger equity stake or agree to supply Reflection with more chips and computing power.

A deal could be reached in the coming weeks, though the people cautioned that the talks could fall apart. The FT could not establish the terms under discussion, but Reflection was last valued at $25 billion in a March funding round. Reflection and Nvidia declined to comment to the FT.

Nvidia is already a major backer

Nvidia has invested $800 million in Reflection and is one of its largest shareholders. Other backers include Sequoia Capital, Disruptive and 1789 Capital, the venture firm where President Donald Trump's son Donald Trump Jr. is a partner.

The Trump administration hopes Reflection can compete with cheaper Chinese open-weight models such as DeepSeek, according to the FT. Reflection has partnered with the Pentagon and the Department of Energy and has agreements to build AI models for U.S. allies including South Korea. Its founders, Misha Laskin and Ioannis Antonoglou, previously worked on Google's Gemini AI model.

Google's recently-announced Gemini 4 Argon model is currently ranked third among all current frontier models in The Latent Score.

Nvidia used the same structure for Groq

Big Tech companies have increasingly used acqui-hires to avoid antitrust scrutiny of AI deals. Nvidia used a similar format in its $20 billion deal with chip designer Groq (unrelated to SpaceX's similar-sounding Grok AI model) in December. That deal drew a lawsuit from Groq employees this week, the FT reported.

Nvidia has its own family of open-weight models, called Nemotron, but no frontier-scale model of its own. Nvidia CEO Jensen Huang called in July for a domestic open-weight "ecosystem," and in September the company agreed to buy model repository Hugging Face for about $13 billion.

Nvidia has also put tens of billions of dollars into AI labs including OpenAI and into cloud providers such as Nebius. Last month it committed $1 billion to U.K. infrastructure firm Nscale as part of a $3.36 billion pre-IPO round.

Beam weights are due this month

Reflection unveiled Beam, its first open-weight model, on Oct. 5. Beam is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active, and was trained on Nvidia GB300 chips, according to Reflection's blog post.

Reflection says Beam matches the Chinese firm Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. Beam scored 65.5 on the SWE Bench Pro coding test against 62.1 for GLM-5.2, but the company's own table shows DeepSeek V4.1 Flash and Kimi K3 ahead of Beam on the DeepSWE and Terminal Bench coding tests.

Reflection plans to release Beam's weights under an Apache 2.0 license later this month and is offering early access through a waitlist for now. France's Mistral also expects to publish weights for its "le Chonk" model by the end of October, The Latent reported last week.