A New York-based AI startup just put a number on America's answer to China's open-source AI surge: 501 billion parameters. Reflection AI unveiled Beam on Monday, a sparse mixture-of-experts model the company says activates just 23 billion of those parameters at a time, built specifically for coding, multi-step reasoning and agentic work. The company is betting that smaller, more efficient compute use can still go toe-to-toe with the Chinese models currently dominating the open-weight race.
A Text-Only Model With Big Ambitions
Beam is Reflection AI's first open-weight model, built as a text-only system, though the company notes it can still process information from other modalities when that information is represented as text, according to Reflection AI. The model was pretrained on 23.8 trillion tokens of what the company describes as diverse, curated, high-quality data. Training the model's reasoning abilities required a serious compute investment too: Reflection deployed 10,500 Nvidia GB300 GPUs for four weeks, generating more than 100 million rollouts during reinforcement learning, per the company's technical materials.
Beam remains in final red-teaming and evaluation, and Reflection says the weights, a technical report, model card and developer artifacts are set to arrive later this month, according to the company's own announcement. Implicator.ai reports those weights are expected under an Apache 2.0 license. Notably, no independent evaluation of Beam had been published as of the outlet's October 5 report, leaving outside verification of the company's benchmark claims an open question.
Benchmarks Claim Efficiency Edge Over Chinese Rivals
Reflection says Beam is competitive with Z.ai's GLM-5.2 on coding and agentic tasks while using three to four times less inference compute, the processing required for a model to answer a prompt, and that it's approaching Alibaba's Qwen 3.8-Max on the same measures, according to the company. The firm also says Beam is nearing comparable scores to GLM-5.2 on advanced reasoning benchmarks while cutting that same compute cost, per Reflection AI's technical blog.
Not every comparison favors Beam. Reflection itself acknowledges that Moonshot AI's Kimi K3 remains ahead on raw capability, framing Beam's selling point as efficiency at inference time rather than outright power. TechCrunch reports that Reflection's benchmarks show Beam outscoring Thinking Machines Lab's Inkling on four coding tests where both models reported results.
Founders With DeepMind Roots and Billions in Backing
Reflection AI was founded in March 2024 by Misha Laskin, now the company's CEO, and Ioannis Antonoglou, both former Google DeepMind researchers, according to Fortune. Laskin has framed the company's open-weight strategy in blunt terms, saying the only way to own intelligence is if it is open. The startup has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, the outlet's report notes, and in June confirmed it closed a funding round at a $25 billion pre-money valuation, as reported by Semafor.
That capital has translated into serious infrastructure commitments. The company signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia's GB300 chips through 2029, per the same TechCrunch report. Reflection AI was reportedly dubbed the “DeepSeek of the West” by some investors, per Business Insider's reporting on a March Wall Street Journal item, a label now circulating widely as Beam enters the conversation — though it's a framing device rather than the full story.
Why This Matters Beyond One Startup's Launch
Open-weight models let developers download the full set of parameters, run them on their own hardware or cloud infrastructure, and fine-tune them for specific use cases — a sharp contrast with OpenAI and Anthropic, which generally provide their models only through their own APIs. That distinction matters because deepSeek, Alibaba, Z.ai and Moonshot AI have all released powerful open models out of China, while on the American side Meta, Mistral, Nvidia and Thinking Machines Lab have released competing open systems of their own.
Despite the attention open-weight launches generate, Axios reports that open-weight models account for only a small proportion of enterprise AI usage overall, even though they sometimes represent a majority of traffic on platforms that serve many different models at once. Some analysts at the Center for Strategic and International Studies estimate open-weight models trail advanced closed frontier systems by roughly four to six months, while still delivering near-frontier performance at a fraction of the cost. The Verge reports that the rise of capable Chinese open-weight models is increasingly pressuring closed-model providers like OpenAI and Anthropic from within their own industry.
Reaction within the AI community was swift. Nvidia's official AI account on X congratulated Reflection on Beam's launch, while investors Deedy Das and Shaun Maguire both expressed enthusiasm for the release on the platform. Axios had reported the model was expected to initially lag behind the most cutting-edge U.S. systems while still competing with top Chinese open-weight offerings — a framing that puts Beam's debut less as an outright victory and more as a marker in an ongoing contest. For now, Fortune notes that Nvidia's own Nemotron model remains the top-ranked U.S.-based system on OpenRouter's leaderboard at No. 6, with OpenAI's gpt-oss trailing at No. 19, underscoring how far American open models still have to climb.
