Reflection AI co-founders Yannis Antonoglou (left) and Misha Laskin. Last August, a Reflection AI co-founder flew to California to meet with NVIDIA CEO Jensen Huang, according to sources. At the time, the startup was developing an AI coding tool, but Huang suggested a pivot toward a more novel path: building open-source AI models that developers could download and modify themselves.
Backed by Huang's strong support and NVIDIA's initial $800 million investment, Reflection CEO Misha Laskin and co-founder and technical lead Yannis Antonoglou positioned the company as a key U.S. counterweight to Chinese players in the open-source AI space. This vision attracted billions of dollars from investors including JPMorgan Chase, Sequoia Capital, and 1789 Capital (a venture firm where Donald Trump Jr. is a partner). The U.S. Department of Defense also added Reflection to its military AI supplier list alongside seven major American tech giants.
Nearly a year after that meeting, however, the global open-source AI boom has exploded, and Reflection has yet to release its own large language model, falling further behind a growing field of competitors. Since DeepSeek launched its R1 model in January 2025, shaking the U.S. tech industry, multiple increasingly powerful Chinese open-source models have attracted significant enterprise customers. This month, Moonshot AI released Kimi K3, which matches top U.S. closed-source models from Anthropic and OpenAI on several tasks. This has also spurred intense U.S. political debate over how to regulate and implement open-source AI.
Domestic competition is equally fierce. Well-funded startups Thinking Machines Lab and Poolside have each released open-source models, and NVIDIA itself has launched its own open-source large model. According to sources close to the company, New York-based Reflection plans to release its first large model this year, with a larger version slated for early 2027 and smaller, lightweight models to follow. The source said Reflection's initial product will match the performance of leading Western models at its release, but will still lag behind current top Chinese open-source models. The company hopes to catch up with subsequent iterations.
"All open-source players are in a fiercely competitive race," said Anastasios Angelopoulos, co-founder and CEO of AI benchmarking startup Arena. Arena's rankings show Kimi K3 topped the charts for UI generation tasks after its release. He believes Reflection can still enter the market if it delivers a quality model, but the sooner it releases an open-source product, the easier it will be to establish a foothold in the crowded space.
Open-source large models can counter the monopoly of closed-source models from OpenAI and Anthropic. As top closed-source models' prices continue to rise, more enterprises are turning to open-source alternatives—not only for lower and more predictable costs, but also for deep customization. Overseas enterprises and governments are increasingly adopting open-source models, and Reflection is betting that many institutions will prefer U.S.-made models over the currently dominant Chinese open-source options.
"DeepSeek R1 was a huge wake-up call for the entire United States," Laskin said in an October interview on TITV, a channel from The Information. "We want to build top-tier open-source intelligence models that rival the best products globally, enabling the U.S. and the world to build various intelligent applications on top of them." While Reflection's first product may not yet achieve this ambitious goal, it still has a ticket to the race. The company has hired a large number of top AI researchers and secured a $25 billion valuation in its latest funding round this year, with investors acknowledging its model training progress. Even if the initial model isn't industry-leading, Reflection can offer customization services to attract customers.
Ambitious vision behind the scenes
Laskin, 36, was born in Russia, moved to Israel as a child, and later settled in the U.S. He graduated from Yale University and earned a PhD in physics from the University of Chicago. In 2022, he joined Google DeepMind, working on early versions of Gemini, and left in 2024 to found Reflection. Co-founder Antonoglou, 38, was born in Greece and earned a PhD in AI from University College London. He spent over a decade at DeepMind, leading the development of AlphaGo, the landmark AI that defeated Go champion Lee Sedol in 2016.
Initially, the duo's startup focused on AI coding tools, a hot area with competitors like Cursor. But after meeting Huang, Reflection almost overnight abandoned its original path. Sources said, "The company transformed from a dual-track development of 'app plus model' to a pure AI lab—a fundamental strategic shift. The two founders completely shelved their original plans and steered the company entirely according to Huang's direction."
Reflection completed a $2 billion funding round in early October last year, with 40% coming from NVIDIA. It raised another $2.5 billion this year. Around the time of the last round, Laskin gathered all employees at a hotel in the Hamptons to rally them toward the new goal of becoming "America's top-tier open-source AI service provider." One employee who attended said the team was inspired and widely believed that with NVIDIA's computing power and financial backing, Reflection could grow into a major industry player.
For Jensen Huang, supporting Reflection helps build a core foundation for the U.S. open-source AI industry, with models deeply optimized for NVIDIA's chips and hardware. With restrictions on sales of hardware to Chinese AI companies and pressure from traditional customers (major U.S. AI giants) developing their own chips, fostering a domestic open-source ecosystem has become a strategic move.
Building a commercial open-source AI ecosystem
After the pivot to a pure AI lab, Reflection set a hiring goal and ultimately built a 150-person research team, according to sources familiar with the NVIDIA meeting. The company now has over 230 employees, with more than 100 researchers. Although no official model has been released yet, the company has begun recruiting a sales team to engage with governments and deploying frontline engineers to customize models for large enterprises.
Reflection has partnered with Dell, which will offer Reflection's models to enterprise customers using its hardware and software, replicating Dell's collaboration with French open-source AI company Mistral AI. The company continues to expand its computing power for model training, recently finalizing agreements to lease NVIDIA AI servers from SpaceX and Nebius, with total costs potentially reaching tens of billions of dollars over the next few years.
Outsize influence in Washington policy circles
Reflection has made a major push into Washington policy circles, hiring Rachel Appleton, Anthropic's first policy lobbying lead. For a company of its size, its influence in federal AI planning far exceeds that of typical peers. In May, eight companies signed contracts with the U.S. Department of Defense to allow the military to use their AI technology on classified networks, with Reflection among them. The other seven were SpaceX, OpenAI, Google, NVIDIA, Microsoft, Amazon Web Services, and Oracle. The U.S. Department of Energy also included Reflection's technology in its "Project Genesis" initiative to accelerate scientific breakthroughs using AI. According to The Information, Reflection has actively participated in White House industry discussions about how federal review rules for frontier AI, outlined in a June executive order, would affect open-source models.
Race on two fast-moving frontiers
While Reflection races to polish its models, competitors are constantly pushing performance boundaries. Before the launch of Kimi K3 this month, the Chinese open-source model Z.ai GLM-5.2 had already garnered widespread developer praise. Cybersecurity tests by the UK government showed GLM-5.2's performance was on par with major U.S. models released just four months earlier.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati in February 2025, released its first open-source model, Inkling, this month. Inkling has nearly 1 trillion parameters, less than half of Kimi K3's parameter count. In the Artificial Analysis intelligence scoring system, Inkling scored 41, Kimi scored 57, and Anthropic's latest model scored 61. NVIDIA updated its own open-source product line in June, launching Nemotron 3 Ultra, with about half the parameters of Inkling and a score of 38, trailing both Inkling and top Chinese open-source models.
To achieve its goal of matching Western open-source leaders, Reflection's first product needs to at least keep pace with Inkling. Sources say Reflection's model has hundreds of billions of parameters, placing it between coding AI company Poolside's open-source model Laguna S 2.1 and Inkling, with a strong emphasis on code capabilities optimized for AI agent operation. "Developing large models is extremely difficult, akin to building a rocket," Laskin said in an April interview. "Small rockets aren't hard to build, but it's tough to make them competitive; building a large rocket that leads the industry inevitably requires a long cycle."
Comments