An open model challenger to Anthropic and OpenAI

Summary · 4 min read

Laskin is Reflection’s CEO and cofounder; he led reward modeling for Google’s Gemini. Antonoglou is Reflection’s CTO and cofounder; he helped create AlphaGo and AlphaZero and led Gemini’s RLHF effort.

Laskin frames open models as a move from renting intelligence to owning and controlling it, while Antonoglou says open models have closed much of the capability gap and adoption should grow as AI markets mature. The founders present Beam, Reflection’s first open-weight model, as a model trained from scratch with 500 billion total parameters and 23 billion active, emphasizing reasoning efficiency and agentic capabilities. They argue openness can improve safety by widening scrutiny, while saying capability thresholds may warrant release restrictions regardless of whether a model is open or closed. They expect to make money by helping enterprises, governments and sovereign customers deploy open models and inference across a broader stack, and say Reflection plans larger models in 2027.

Minute by minute

1. 0:41

   AI adoption is shifting toward ownership. Laskin compares closed models to renting intelligence and open models to owning it: users can run, control and customize models on their own infrastructure. He says adoption accelerated as startups grew into enterprises; on OpenRouter, he estimates the split flipped from 30% open and 70% closed six months earlier to 70% open and 30% closed.

2. 5:53

   Laskin says geopolitical support helped China lead. Laskin says Western commercial incentives initially favored closed models because AI was a rental market. He attributes China’s push for open models primarily to geopolitical motives, and says DeepSeek V3’s performance raised the relevance of Chinese models as a commercial market for open models emerged.

3. 8:00

   Beam targets efficient agentic work. Laskin says Beam is Reflection’s first open-weight model, with around 500 billion total parameters and 23 billion active. He highlights its coding and agentic capabilities and enterprise and public-sector utility. Antonoglou says it is three or four times more token-efficient than some open models, including GLM 5.3, and some closed models, including GPT-6 Luna.

4. 14:01

   Training from scratch supports reinforcement learning. Antonoglou says Reflection trained Beam from scratch to build a Western frontier model with its own team and to control pretraining. He says the team optimized data, architecture, router balance and learning dynamics end to end so the mixture-of-experts model would remain stable during large reinforcement-learning runs.

5. 15:56

   Laskin argues openness can improve safety. Laskin compares AI to software and says open protocols and wider inspection can expose vulnerabilities that a small group of closed-lab researchers may miss. Antonoglou says sharing with researchers has helped produce safety insights. The founders say highly capable models could require gated access or regulation based on capability, regardless of whether they are open or closed.

6. 23:51

   Open models add safety options. Laskin says open and closed models can provide countervailing tools; he cites Hugging Face turning to open models after closed-model safeguards blocked its cyber-defense use. Antonoglou says Reflection evaluates both risk and usefulness, then uses examples, supervised fine-tuning and reinforcement learning to shape model behavior.

7. 32:25

   Reflection began with two research bets. Laskin says the founders started Reflection after working on Gemini, betting that reinforcement learning could produce agentic coding capabilities and that Western open models would make research more cost-effective. He says the company showed early reinforcement-learning results, then changed course as the Western open-model frontier receded after DeepSeek V3.

8. 37:23

   Open models can drive services revenue. Laskin says open models can drive demand for cluster management, serving, inference and agent software rather than being the revenue product by themselves. Reflection aims to help enterprises, public-sector groups and sovereign customers deploy and operate those systems. Antonoglou describes AI as labor that large institutions will want to control.

9. 42:19

   Trust, compute and models shape revenue. Laskin says owning more of the AI stack could let Reflection capture margin, but argues the near-term priority is strong models and more inference. He describes the business opportunity as intelligence density multiplied by compute under the company’s umbrella and customer trust, and says building models helps win enterprise trust.

10. 46:30

    Domestic open infrastructure matters. Antonoglou says the provider of default open AI infrastructure can shape the economy built on it, as open software did in the internet era. He says using models that customers can inspect and obtain from a trusted provider may also reduce security risks as agents gain the ability to act on harmful prompts.

11. 48:39

    Partnerships distribute open models. Laskin says Reflection will build tools and work with cloud providers, hyperscalers and hardware partners to integrate its models across the stack. He describes a 250-megawatt data center in Korea and Dell’s role in infrastructure. He says Reflection has built cluster-management, inference and agent-sandboxing software through its own model training and can provide it to customers.

12. 54:41

    Laskin expects a multipolar AI market. Laskin argues that open and closed providers can coexist because AI markets will be broad and demand for artificial labor will grow. He favors a multipolar ecosystem, saying open models let other companies build businesses on top of model providers and that 10 or 20 intelligence providers would be better than two or three.

13. 56:40

    Reflection plans larger models in 2027. Antonoglou says Reflection plans significantly larger models in 2027, followed by repeated releases as reinforcement learning scales in waves. Laskin says the team ran what it believes is the largest openly documented reinforcement-learning run, with around 10,000 GB300s running for weeks, and says the model continued learning as more compute was applied.

Published Oct 5, 2026.