Two years and $4.6 billion in funding later, a startup called Reflection is soon releasing its first open-weight model in a bid to become the DeepSeek of the West.
Reflection’s new model, Beam, is arriving at an interesting time: Chinese labs have been setting the pace at the top end of open models for a while, and the Western world has lacked an answer. As more workloads shift to open models, who develops them matters more. A competitive Western alternative to Chinese open models would give businesses and governments another serious option for AI they can run and customize themselves.
Reflection’s founders have helped build some of the systems that got us here. CEO Misha Laskin led reward modeling for Gemini at Google DeepMind. CTO Ioannis Antonoglou helped create AlphaGo and AlphaZero, then led Gemini’s reinforcement learning from human feedback effort. Nvidia, Sequoia, and Lightspeed are among the investors backing Reflection, which was last valued at $25 billion.
Beam was trained from scratch and has 501 billion parameters. Reflection says early tests put it alongside leading open models on coding and agent tasks, and that it uses roughly a quarter of GLM-5.3’s inference compute in its reasoning-efficiency comparison. An Apache 2.0 license also means developers can modify it and build businesses on it without special commercial restrictions. “Closed models are the equivalent in real estate to renting an apartment,” Laskin says. Now, he argues that “an ownership market” for AI is starting to form.
The big question is how Reflection makes money while giving Beam away. We get into that, its plans to help companies and countries run their own AI, and why the founders think a handful of closed labs shouldn’t control where the technology goes. I also press them on safety and whether they'd consider a future model too powerful to release.
Sources is available wherever you get podcasts, including video or audio on YouTube and Spotify, and audio on Apple Podcasts.
Thanks to the show’s premiere sponsors: Atlassian, Granola, and Mercury.
A MESSAGE FROM MY SPONSOR
Powering Innovation
Constellation is America’s largest producer of reliable, clean energy, driving progress and innovation coast to coast. From local neighborhoods to iconic cities, we bring the energy that serves industries, strengthens communities, and moves America forward.
Episode highlights
Reflection trained Beam from scratch, without distilling another lab’s model, because it wanted control over the entire training process. Antonoglou says owning the architecture, data, and pretraining was essential to pushing reinforcement learning further. The initial training had to give Beam the reasoning capabilities that reinforcement learning could then build on. That control also helped the team keep the model stable as it poured more compute into training. There was also a geographic consideration: Reflection wanted its teams in the US and UK to develop a Western model end-to-end.
Reflection’s business is selling the software and infrastructure that make its open models useful to companies and governments. Customers need help deploying agents, customizing models, and running them at scale, which is where Reflection expects to make money. Its $1 billion compute deal with Nebius and multibillion-dollar agreement with SpaceX give a sense of the scale of that bet. Over the long term, Laskin says it “probably makes sense” for Reflection to become its own hyperscaler, allowing Reflection to capture more margin. For now, the priority is getting its models out and building demand for running them.
Reflection’s founders see concentrating AI in a handful of closed labs as a safety risk. Laskin argues that a small group of researchers cannot find every vulnerability, while open models let a much larger community inspect the technology and develop safeguards. Antonoglou says putting so much power inside one or two institutions creates “a single node of failure,” even when those institutions have good intentions. Their preferred outcome is a mix of open and closed systems, with different providers able to help defend against failures elsewhere. When I press them on whether a future Reflection model could be too powerful to release freely, they acknowledge that greater capabilities could justify restrictions. Laskin says those decisions should turn on what a model can do: “That is independent of whether they are open or not.”
ICYMI
Sources is a newsletter and podcast by Alex Heath about the AI race, featuring conversations with leading founders and technologists. Every week, Sources reaches the inboxes of thousands of decision-makers in tech, finance, policy, and media.
