TL;DR
Reflection AI's Beam, a 501B-parameter open-weight model from two ex-DeepMind founders, launches as the most credible US answer yet to China's open-model surge, arriving the same week DeepSeek dropped a rival efficiency model and as evidence mounts that Chinese military researchers are distilling American AI into defense systems.
What happened
- Reflection AI unveiled Beam on October 6, pitching it as a US open-weight frontier model for coding, reasoning, and agentic tasks.
- 501 billion total parameters, 23 billion activated per token: a mixture-of-experts design built for cheap, fast inference.
- Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers, and raised at a $25 billion pre-money valuation in June, backed by Nvidia, Sequoia, and Citigroup.
- Beam is currently on a waitlist during final safety testing; weights, a technical report, and developer tools are due later in October.
- Simultaneously, DeepSeek released V4-Flash, a 284B-parameter MIT-licensed model trained on 32 trillion tokens, priced at $0.14 per million input tokens and benchmarking close to Anthropic's Claude Opus 4.8 on several agent tasks.
Why it matters
- The open-weight race is now a geopolitical contest: Beam is explicitly positioned as a homegrown alternative to DeepSeek, Alibaba Qwen, Z.ai GLM, and Moonshot Kimi, all of which have outpaced US open models in recent months.
- Reflection claims Beam matches Z.ai GLM-5.2 on coding and agentic benchmarks using 3 to 4 times less inference compute, a direct efficiency challenge to Chinese labs.
- DeepSeek V4-Flash's simultaneous release shows the competitive pressure is real: MIT-licensed, self-hostable, and already beating GLM-5.2 across all nine published agent benchmarks.
- A Jamestown Foundation review of Chinese academic papers found PLA-affiliated researchers using OpenAI and Anthropic outputs via adversarial distillation to train military and surveillance AI, with the White House confirming Kimi K3 was distilled from Claude Fable, raising the stakes for every open-weight release.
- Sovereignty framing is the new sales pitch: Laskin's quote, "the only way to own intelligence is if it's open," signals that US open-weight models will compete on control and trust, not just benchmark scores.
What to watch next
- Whether Beam's full weight release later in October holds up on independent benchmarks, especially against DeepSeek V4-Flash and Kimi K3, which Reflection concedes is still ahead on raw capability.
- US regulatory response to the adversarial distillation findings: the White House has already named Kimi K3, and export controls or API access restrictions on closed models could reshape how Chinese labs build the next generation.
- Reflection's path to revenue: at a $25 billion valuation with no product yet shipping, the company needs enterprise adoption of Beam to justify the number before a larger funding round or IPO.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.