Reflection AI
Reflection AI is an artificial intelligence company focused on open models, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou; it initially worked on agent tools for AI-assisted software development. On August 9, 2025, the company announced on its blog its pivot to “frontier open intelligence” and said it had raised $2 billion.
Openness commitments and business
The official About page states three commitments: releasing open weights for its models, publishing the research behind them as papers and technical reports, and open-sourcing software that lets the community customize and build on the models (including reinforcement learning tools and environments). The business spans the open-weight models themselves, deployment infrastructure (an API platform or self-hosted deployment), and a forward-deployed team that works with enterprises.
Collaborations listed on the About page include validation on Dell AI Factory with NVIDIA, providing open-weight models to the U.S. Department of Energy’s Genesis Mission (serving the 17 U.S. National Laboratories), and building a sovereign AI cloud in South Korea with Shinsegae, powered by its open-weight models and NVIDIA GPUs.
The Beam model. On October 5, 2026, the company announced on its blog Beam, its first open-weight model: a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, trained with a focus on coding, reasoning, and agentic workloads. The post includes benchmark comparisons against open models such as GLM 5.2, Kimi K3, and Qwen 3.8 Max, and says the weights, technical report, and model card would be released later that month. Beam’s current release and availability status should be read from the official site.
Discussion in the show
In Weekly #005’s chapter “开源模型与 AI 意识的讨论” (open models and AI consciousness), discussing newly released open models, Yang Pan refers to “Reflection, the one NVIDIA invested in” (paragraph in the Chinese transcript) and offers a personal comparison placing it around GLM 5.2’s level and behind the Chinese labs. Beam’s official blog does use GLM 5.2 as one of its benchmark references, but the company makes no “behind” claim of that kind; the details of NVIDIA’s investment stake have not been verified item-by-item in official channels, so the show’s phrasing is not treated as fact here.
Frequently asked questions
Who founded Reflection AI?
Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, per Wikipedia; the team’s background (contributors to PaLM, Gemini, AlphaGo, AlphaCode, and more) is described on the official blog.
What does Reflection AI do?
Reflection AI develops open-weight foundation models plus the open-source software and deployment infrastructure around them, committing to open weights, published research, and open-source software, as stated on the official About page.
What is Reflection AI’s Beam model?
Beam is the first open-weight model announced by Reflection AI, on October 5, 2026: a sparse MoE architecture with 501 billion total parameters and 23 billion active, focused on coding, reasoning, and agentic workloads, per the announcement post. Its current release status should be read from the official site.
How much has Reflection AI raised?
The company’s blog post of August 9, 2025 announced $2 billion raised for building “frontier open intelligence,” per that post. Later valuation reports have not been verified against primary official text and are not cited here.