Reflection’s founders on building a DeepSeek of the West

October 5, 2026

AI Summary

5 min read

Reflection’s founders on building a DeepSeek of the West

Misha and Giannis met at DeepMind working on Gemini 1.5, where Giannis was the RLHF lead and Misha was on his team. When they founded Reflection in 2024, they held two bets: that reinforcement learning was the path to agentic and coding capabilities, and that powerful Western open models would exist to build on. The first bet proved right quickly. The second didn't. After DeepSeek V3 landed, the Western open frontier receded, and no equivalent appeared for months. So Reflection changed course: instead of building on top of someone else's open model, they would build the open model itself. The result is Beam, a roughly 500-billion-parameter mixture-of-experts model with 23 billion active parameters, trained from scratch in the United States and UK.

Why open models are suddenly ascendant

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Sources with Alex Heath

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:08) **The Open Model Thesis** - Misha introduces the "rental vs. ownership" real estate analogy to explain the surge in open model demand.
2 (03:40) **Why Open Models Keep Pace** - Giannis argues there is no technical reason open models cannot match closed frontier performance.
3 (06:02) **Why China Won Open (So Far)** - Misha explains the geopolitical and market incentives that drove Chinese labs to dominate open models.
4 (08:16) **Introducing Beam: The First Reflection Model** - Misha and Giannis unveil their 500B parameter, 23B active parameter open-weight model.
5 (12:51) **Consumer Applications and Ecosystem** - The founders explain how a compact, efficient model enables broader consumer use cases.
6 (14:10) **Training From Scratch vs. Distillation** - Giannis explains why Reflection chose to pre-train Beam entirely from scratch.
7 (16:04) **Safety and Alignment in Open Models** - Misha argues that open models are fundamentally safer than closed models, using the "white blood cell" analogy.

+ Full timestamped outline available in the app

Show Notes

Reflection is soon releasing Beam, its first open-weight AI model, in a bid to become the DeepSeek of the West. Trained from scratch, Beam is designed for coding, reasoning, and AI agents. Reflection's benchmarks put it alongside the strongest open models, with substantially more efficient token economics.

To mark Beam’s debut, I'm joined by Reflection's co-founders: CEO Misha Laskin, who led reward modeling for Google’s Gemini, and CTO Ioannis Antonoglou, who helped create AlphaGo and AlphaZero before leading Gemini’s RLHF effort. Reflection has raised $4.6 billion from investors that include Nvidia, Sequoia, and Lightspeed.

We discuss why Chinese labs pulled ahead in open AI models, how Reflection built Beam, and why the founders think companies and countries will increasingly want to control their own models. I ask how they plan to make money from open-weight AI, whether openness can make these systems safer, and what comes next for Reflection.

Thanks to the show’s premier sponsors: Atlassian, Granola, and Mercury.