Nvidia-Backed Reflection AI Unveils 501-Billion-Parameter Beam Model as Unconfirmed Buyout Talk Circulates

Since Reflection AI announced Beam on October 5, the Nvidia-backed startup has gone from a company with no shipped product to the American challenger to China's open-weight models. Now a post circulating on X claims Nvidia is in talks to buy it outright. Nvidia has not confirmed any such discussions, and nothing has substantiated the claim.

Nvidia already has roughly $800 million in the company. Reflection has raised more than $4 billion in total, with other backers including Sequoia Capital, Lightspeed, Disruptive and 1789 Capital, which counts Donald Trump Jr. on its board. Its latest round set a pre-money valuation of $25 billion, roughly triple the $8 billion of an earlier round.

What Beam is

Beam is a mixture-of-experts model with 501 billion total parameters. Reflection says only about 23 billion, around 4.6%, activate for any given task. The company says it trained the model on 23.8 trillion tokens using 6,144 Nvidia GB300 NVL72 GPUs, and that it has a 1 million token context window.

The full weights are scheduled to be released later this month under an Apache 2.0 license. Anyone will be able to download, modify and run the model on their own hardware.

The founders are Misha Laskin, who led reward modeling on Google DeepMind's Gemini project, and Ioannis Antonoglou, who co-created AlphaGo. Laskin has said the West needs its own answer to Chinese open models.

The numbers are Reflection's own

Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning while using three to four times less compute at inference. It reports 80.9 on SWE-Bench Verified and 97.8 on AIME 2026. It says Beam is near Alibaba's Qwen 3.8 Max for coding.

The company's own comparison tables show a mixed picture, and Beam trails Qwen 3.8 Max rather than leading it. On DeepSWE v1.1, Beam scored 44.4 against 51.0 for Qwen 3.8 Max. On Terminal Bench v2.1, Beam posted 80.1 versus 86.6 for Qwen. Moonshot AI's Kimi K3 remains the most capable of the group.

Artificial Analysis, the benchmarking platform, says it had early access to Beam and that the model is on track to be among the most efficient at its level of intelligence. Independent results on the public release are still to come.

Compute commitments

Reflection has lined up infrastructure well ahead of its product. It has compute arrangements worth up to $6.3 billion with SpaceX, described as a potential partnership. It also has a deal worth more than $1 billion with Nebius for Nvidia GPU access through 2029.

The company also plans a 250 MW AI data center in South Korea with Shinsegae Group, to run on Nvidia technology. Reflection also works with Nvidia through the Nemotron Coalition, Nvidia's open-model initiative.

The commercial pitch is what Reflection calls the "AI factory." Enterprises and governments would combine their own data, Reflection's models and their own compute to run localized systems without sending data back to a third-party lab. Axios reported that Reflection has held discussions with interested parties in Washington and elsewhere to detail the release and explain how the AI factory would work. The concept echoes a vision Nvidia CEO Jensen Huang has described for years.

The China problem and the policy shift

The strongest open-weight models today come almost entirely from China, including DeepSeek, Alibaba's Qwen, Moonshot's Kimi and Z.ai's GLM. DeepSeek's R1 release in January 2025 showed that a frontier-grade model could be trained and given away at far lower cost than U.S. labs had budgeted. Meta, the most prominent Western open-model developer, has not matched the leading Chinese releases, per its recent results.

Months ago, Nvidia, Meta, Microsoft and Palantir signed a public letter urging the U.S. government to support American open-weight development rather than restrict it. The White House called for similar measures weeks later.

There is a counterargument. Axios noted that open models can be harder to monitor and regulate than the systems from OpenAI, Anthropic and Google. Backers of open ecosystems answer that wide availability and transparency bring their own security advantages.

What is and isn't known about a deal

The acquisition claim comes from a single post on X. Nvidia has not confirmed any such discussions, and the talks have not been independently substantiated. Nvidia already funds Reflection, supplies its GPUs and works with it through the Nemotron Coalition, so the relationship is close whatever happens.

No price, structure or timetable has surfaced. Reflection's last round set a $25 billion pre-money valuation, a steep tag even for a company with Nvidia already on board. A purchase would also raise a strategic tension: Nvidia sells chips to many AI labs, and owning a model developer could make it a competitor to some of its own customers.

The next checkpoint is the Apache 2.0 weight release later this month. Outside researchers will then be able to test Beam against Reflection's claims on the efficiency and coding scores.