TheChek News Flow

The Takeaway

Last August, a co-founder of Reflection AI flew to California to meet Nvidia CEO Jensen Huang. The fledgling startup had been making an AI coding tool, but the tech magnate encouraged him to go with a new plan, said someone familiar with the meeting, for something that at the time seemed more novel: developing open-source AI models that developers can download and modify.

With Huang’s backing and $800 million in initial investment from Nvidia, Reflection CEO Misha Laskin and his co-founder and technology chief, Ioannis Antonoglou, positioned their company as America’s champion in a contest with Chinese firms developing open-source AI. That vision has drawn billions of dollars of capital from other investors, including JPMorgan Chase, Sequoia Capital and 1789 Capital—a venture firm at which Donald Trump Jr. is a partner—and led to its inclusion by the Pentagon in a group of AI suppliers along with seven of the biggest tech firms in America.

But as interest in open-source AI crescendos nearly a year after the meeting with Huang, Reflection has yet to actually release its AI model—and it is falling further behind a growing list of rivals.

Increasingly powerful models from China have gained traction with businesses since the launch of DeepSeek’s R1 model first jolted the U.S. tech sector in January 2025. Moonshot AI’s Kimi K3, released this month, performed as well on some tasks as top U.S. models from Anthropic and OpenAI. That has fueled discussion among U.S. officials about whether and how to regulate open-source AI. Domestic competition also has increased with recent open model releases from well-funded startups Thinking Machines Lab and Poolside—and Nvidia itself.

Reflection, based in New York, is planning to release its first model this year, with a larger version slotted for early 2027, followed by a smaller version, according to someone close to the company. Reflection’s first model will be on par with the best Western models at the time of its release, this person said, meaning it will lag behind the leading Chinese open models, though Reflection aspires to exceed those with future releases.

“For all the open-source providers, it’s an extraordinarily competitive area,” said Anastasios Angelopoulos, co-founder and CEO of Arena, an AI evaluation startup that crowned Kimi K3 as the top model for creating user interfaces after its release. He said Reflection could still jump into the race with a great model, but the sooner it can release an open-source model, the better its chances of getting a foothold in this increasingly competitive market.

Open models provide a counterweight to proprietary ones from providers like OpenAI and Anthropic. As prices soar for the leading proprietary models, more companies are turning to open source, which can offer cheaper and more predictable pricing as well as greater customization.

As more foreign companies and governments adopt open-source models, Reflection is betting some customers will prefer using a U.S. model over the Chinese alternatives that so far dominate.

DeepSeek R1’s arrival was “a massive wake-up call for us in the U.S.,” Laskin said in an interview with The Information’s TITV in October. “We are building the frontier open model intelligence that will punch with the best of them and will be the model that both the states and the world externally can build on top of for their intelligence needs.”

While Reflection won’t meet that ambitious goal with its first model, it is still in the race. The company, which declined to make Laskin available for interviews for this article, has assembled top AI researchers and says progress in training its AI helped fuel the $25 billion valuation it attained in its latest fundraising round earlier this year. Though Reflection’s forthcoming model won’t be cutting edge, the company could customize it to encourage customers to use it.

Grand Ambitions

Laskin, 36, was born in Russia and moved as a child to Israel and then to the U.S., where he graduated from Yale and earned a doctorate in physics at the University of Chicago. He joined Google’s DeepMind AI unit in 2022 and contributed to the early versions of Gemini before leaving to start Reflection in 2024.

His co-founder, Antonoglou, 38, was born in Greece and earned his doctorate in AI at University College London. During more than a decade at DeepMind, he helped develop AlphaGo, the pioneering AI system that defeated Go champion Lee Sedol in 2016.

The pair started by developing an AI system for coding, then a buzzy area thanks to the rise of startups such as Cursor. But Reflection dropped that plan essentially overnight after the meeting with Huang.

“We went from a product-plus-model builder to only a lab, and that was the big pivot,” said the person with knowledge of the meeting. “The co-founders are completely 100% steering the company away from their vision to Jensen’s vision.”

Reflection raised $2 billion in early October—40% from Nvidia—and another $2.5 billion in this year’s round.

Around the time of the earlier fundraise, Laskin assembled his employees at a hotel in the Hamptons to rally them around the new vision to become the premier provider of open-source AI models in the U.S. His vision was well received, according to one employee who heard his remarks and who said they felt convinced that Reflection would have access to the funding and computing resources necessary to become a serious player in the industry, thanks in part to Nvidia’s support.

For Huang, Reflection at the time represented an opportunity to form the backbone of a U.S. open-source AI movement that could build on Nvidia’s chips and other equipment. Nvidia has faced restrictions selling that gear to Chinese AI firms, and is contending with competition from big U.S. AI companies that have been its clients but are increasingly developing their own AI chips.

Building an Open-Source AI Business

To become a full-fledged AI lab, Reflection set a new recruiting goal to have 150 researchers on its staff, according to the person with knowledge of Reflection’s Nvidia meeting. Reflection now numbers over 230 employees, over 100 of whom are researchers, according to the person close to the company.

Despite not yet having a model, it is hiring salespeople to target global governments and forward-deployed engineers to customize its models for large enterprises. The startup has struck a partnership with Dell to distribute its models to enterprise customers that use Dell’s AI hardware and software—mirroring a deal the computer giant has with Mistral AI, a French open-source company.

Reflection also has been lining up more of the computing power needed to train its AI models, including deals in recent weeks to rent Nvidia AI servers from SpaceX and Nebius that could cost Reflection billions of dollars over the next several years.

Punching Above Its Weight in Washington

Reflection has quickly built a substantial presence in Washington—with hires including Anthropic’s first lobbyist, Rachel Appleton—and has played an outsize role in federal AI plans for a company its size.

Reflection was one of eight companies to sign deals with the Defense Department in May that allowed the military to use their AI on classified networks. The others were SpaceX, OpenAI, Google, Nvidia, Microsoft, Amazon Web Services and Oracle. The Department of Energy separately agreed to incorporate Reflection’s technology in its Genesis Mission, an effort to use AI to accelerate scientific discovery.

And Reflection has actively participated in industry discussions with the White House about how plans for federal review of frontier AI models set up in a June executive order will affect open-source models, The Information has reported.

A Tale of Two Frontiers

As Reflection races to build its AI models, competitors keep pushing the cutting edge further ahead. This month’s release of Kimi K3 came weeks after another new Chinese open-source model, Z.ai’s GLM-5.2, became an instant favorite among developers. In testing of its cybercapabilities, the U.K. government found that GLM-5.2 rivals U.S. models released as little as four months before it.

Thinking Machines Lab, launched in February 2025 by Mira Murati, OpenAI’s former chief technology officer, released its first open-source model, Inkling, this month. At just under 1 trillion parameters—the numbers that make up an AI model and determine its behavior—Inkling is well under half the size of Kimi K3. Inkling earned a 41 score on an index of intelligence from benchmarking firm Artificial Analysis, compared to 57 for Kimi and 61 for Anthropic’s newest model.

Nvidia extended its own line of open-source models in June, with the release of Nemotron 3 Ultra, which is a bit over half Inkling’s size and earned a score of 38 from Artificial Analysis, placing it below Inkling and leading Chinese models.

For Reflection’s first model to perform on par with the best open Western models, it will have to at least rival Inkling. Reflection plans for the model to have hundreds of billions of parameters: That’s between the size of Laguna S 2.1, the new open-source model from AI coding startup Poolside, and Inkling. Reflection’s model will have strong coding capabilities and will be designed to power AI agents, said the person close to the company.

“These things are hard to build,” Laskin said in an interview with CNBC in April. “They’re kind of like rocket ships. And so you can build a small rocket ship, but that’s not really going to impress anyone. To build a big rocket ship, it takes time.”