TL;DR
Reflection AI, a $25B-valued New York startup, has released Beam, a 501B-parameter open-weight model that matches leading Chinese open models at 3 to 4x lower inference cost, framing it as the US answer to DeepSeek.
What happened
- Reflection AI unveiled Beam on October 6, 2026, its first frontier open-weight model targeting coding, reasoning, and agentic tasks.
- 501 billion total parameters, but only 23 billion activate per inference, a mixture-of-experts design that slashes compute costs without sacrificing benchmark performance.
- Weights, a technical report, and developer tools are not yet public; Beam is in final safety testing with access via waitlist, full release expected later in October.
- Founded in 2024 by former Google DeepMind researchers Misha Laskin (CEO) and Ioannis Antonoglou, the company has raised billions and closed a funding round at a $25 billion pre-money valuation in June 2026.
- Backers include Nvidia, Sequoia, and Citigroup; Nvidia AI's official account publicly congratulated the launch.
Why it matters
- Efficiency is the competitive wedge: Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3 to 4x less inference compute, a direct challenge to Chinese open-model dominance on cost grounds.
- Open-weight positioning targets enterprise sovereignty: developers can download, run, and fine-tune on their own hardware, bypassing OpenAI and Anthropic APIs entirely and giving organizations direct control over their AI stack.
- The "DeepSeek of the West" narrative (first reported by the Wall Street Journal in March 2026) gives Beam political and commercial tailwinds as US policymakers and enterprises seek domestic alternatives to Chinese frontier models.
- Crowded field raises the bar: Beam competes not only with DeepSeek, Alibaba Qwen, and Moonshot AI's Kimi K3, but also with open models from Meta, Mistral, Nvidia, and Thinking Machines Lab, meaning benchmark claims will face rapid independent scrutiny.
- Kimi K3 still leads on raw capability by Reflection's own admission, leaving a clear performance gap the company must close to sustain its valuation.
What to watch next
- Full weight release later in October: independent evals from the research community will either validate or deflate the efficiency and benchmark claims.
- Enterprise adoption signals: whether Fortune 500 or government customers move workloads to Beam would confirm the sovereignty pitch is landing beyond investor hype.
- China model response: Alibaba, DeepSeek, and Moonshot AI have all shipped rapidly iterating releases; a counter-update from any of them could reset the competitive benchmark before Beam's weights are even public.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.