Scoop: Powerful open model is set to shake up AI race

Open models, designed to compete with Chinese rivals, can be harder to monitor and regulate than the Big Three so-called frontier models. Supporters of such ecosystems say their widespread availability and transparency also bring security advantages.

• Reflection's first model is set to be joined by other open-weight models from other Western players this month, sources tell Axios. That will bring a new competitive dimension to the AI boom.

State of play: Reflection's model, set to be released soon, is expected to initially lag behind the most cutting-edge U.S. models and be competitive with top Chinese open-weight models.

• Sources tell Axios it will still boast powerful intelligence — capable enough to help companies build their own proprietary, low-cost AI systems.

• Combining a lower-tier open model with high-quality company data can yield results that rival the most expensive frontier AI systems in certain situations.
• A Reflection spokesperson declined to comment.

The big picture: Open-weight models, the most powerful of which are almost uniformly made in China, have taken Silicon Valley by storm.

• As opposed to closed models like Claude and ChatGPT, open-weight models like Reflection's are generally available to be downloaded freely and can be used and customized by anyone.
• On platforms that serve many models, their combined market share at times has surged to take up a majority of usage.
• However, in the far higher-spending arena of enterprise use, which is typically handled through API and corporate billing, open-weight models account for a small proportion of usage, AI execs and analysts say.

Zoom in: Reflection's ultimate goal is to pioneer the "AI factory" — a product allowing institutions to spin up their own localized AI ecosystems.

• A company, for example, would take its own proprietary data, use Reflection's models and secure its own computing firepower to build a highly customized, cheap AI system.
• Some institutions, particularly hedge funds and trading firms, are already eager to build proprietary AI systems using their own highly guarded data.
• The startup has already begun testing this concept, announcing a sovereign AI factory partnership with Shinsegae Group in South Korea.

Behind the scenes: In preparation for the launch, Reflection has held discussions with interested parties in Washington and elsewhere in recent weeks to detail its model release and explain how the AI factory will work, sources told Axios.

• It is also aggressively locking up the compute capacity required to scale. In recent weeks, the startup signed massive deals with Nebius and SpaceX to rent Nvidia AI servers.

What they're saying: Reflection CEO Misha Laskin has indicated it will take time for its models to reach the most advanced capabilities.

• "They're kind of like rocket ships," Laskin, who previously worked at Google, told CNBC. "To build a big rocket ship, it takes time."

The Nvidia factor: The AI factory has been a core vision of Nvidia CEO Jensen Huang for years. He has described a scenario combining computing hardware with open-source models, allowing enterprises to retain ownership of their data.

• Top AI users in Western business and government, such as banks or the Pentagon, balk at using highly capable Chinese models due to security risks. Reflection and other Western players are trying to fill that void.
• For years, Huang has sought to strengthen the open AI ecosystem and has been strongly supportive of Reflection's move toward building a top open system, according to stories in The Wall Street Journal and The Information.

The bottom line: Reflection's release, and the startup's close ties to Nvidia, will bring a renewed focus on open-weight models, combining with other coming offerings to shake up the AI race.