Nvidia Is Reportedly Weighing a Takeover of Reflection AI, Whose 501B Beam Model Has No Public Weights Yet

Nvidia is in early-stage talks to acquire Reflection AI or deepen its investment in the open-model startup, the Financial Times reported on Saturday, October 10. Neither company has announced a deal, and Reuters said it could not verify the report.

The FT said it relied on people with direct knowledge of the matter, according to Reuters. Reuters said Nvidia and Reflection did not respond to requests for comment sent outside business hours.

What the FT reported

Three structures are under discussion, according to Investing.com's summary of the FT: a full acquisition, more financial backing, or an acqui-hire. In an acqui-hire Nvidia would hire Reflection's staff and license its technology, which the FT said could avoid a lengthy regulatory review.

An agreement could come within weeks, the FT said, but the talks could fall apart. Nvidia has already invested $800 million in Reflection, per the FT.

Reuters recalled that Chief Executive Misha Laskin told CNBC in April that Reflection was raising money at a $25 billion pre-money valuation. Reflection was founded in 2024 by Laskin and Ioannis Antonoglou, both formerly of Google DeepMind.

Beam has a licence promise and no weights

The timing follows Reflection's October 5 announcement of Beam, its first model. It is a sparse mixture-of-experts text model with 501 billion total parameters and 23 billion active per token, pretrained on 23.8 trillion tokens.

Reflection said pretraining used 6,144 Nvidia GB300 NVL72 GPUs for under four weeks. A four-week reinforcement learning run used 10,500 GB300s and generated more than 100 million rollouts, which the post called "one of the largest scale RL runs conducted by any open lab to date".

The licence is Apache 2.0, but the weights are not out. Reflection said it will release them, a technical report and a model card "later this month". Until then access is a waitlist for a select group.

Context reached 256,000 tokens during reinforcement learning, and the post says midtraining extends the effective window to 1 million.

Where Beam sits on Reflection's own table

Every figure below is vendor-supplied. The post says it updated Beam's results on October 8, and the table comes from that version.

Benchmark	Beam	Qwen 3.8 Max	Kimi K3
SWE Bench Pro v1	65.5	67.7	not reported
Terminal Bench v2.1	80.1	86.6	88.3
DeepSWE v1.1	44.4	51.0	68.0
GPQA Diamond	90.5	92.6	93.5

Beam beats Nvidia's Nemotron 3 Ultra on every row where both report, including 80.9 against 70.7 on SWE Bench Verified. It trails Kimi K3 on all three rows above where Kimi reports. On DeepSWE v1.1 the table also lists DeepSeek V4.1 Flash at 74.2 to Beam's 44.4.

1. DeepSeek V4.1 Flash90.6 score
2. Kimi K388.3 score
3. Qwen 3.8 Max86.6 score
4. GLM 5.281 score
5. Beam80.1 score
6. Nemotron 3 Ultra56.4 score

Source: Reflection AI, Introducing Beam, accessed 2026-10-11

VKTR reported that Artificial Analysis was given early access and is benchmarking Beam independently, with results not yet published. Alibaba's Qwen3.8 27B is the open-weight comparison point the site has covered, and Cloudflare's Clef is another open release.

Reflection's weights are due before October 31. Nvidia has set no date for a decision on the talks.