Reflection AI unveils Beam as a Western open-weight rival to Chinese models
CEO Misha Laskin is pitching the model to companies and governments that want control over where AI runs; Reflection says Beam's weights will arrive later in October.
By Ryan Merket · Published
Primary source: Semafor
Why it matters
Beam puts Misha Laskin's open-model thesis into a product aimed at institutions that want control over deployment. Its eventual weights and technical details will let developers test whether Reflection's capability and cost claims translate into a usable alternative to Chinese open models.
Reflection AI unveiled Beam on October 5th, its first frontier open-weight model, with CEO Misha Laskin (@MishaLaskin) framing the launch around a practical problem for businesses and governments: they want advanced AI they can run and control themselves, and today's strongest open models largely come from Chinese labs.
Laskin told Semafor that customers looking to build sovereign AI systems "don't really have very good options today." Beam is Reflection's answer to that gap, aimed at reasoning, coding and agentic tasks. Its intended buyers include enterprises and public-sector organizations that want to customize models and run them on their own infrastructure.
The choice of problem fits Laskin's path into AI. He earned a PhD in theoretical physics at the University of Chicago, did postdoctoral research at UC Berkeley and worked at Google DeepMind, where he says he contributed to Gemini. Before founding Reflection, he also started a Y Combinator-backed company. His co-founder, Ioannis Alexandros Antonoglou (@real_ioannis), brought a different part of DeepMind's history: he worked on reinforcement-learning systems including AlphaGo and AlphaZero. Together they are betting that experience building research systems can be turned into models institutions can operate for themselves.
That bet now has a model attached to it, though the public release is still staged. Reflection says Beam's weights and full technical details will be released later in October, with distribution through cloud providers and integrations with open-source libraries. Until those weights and details arrive, developers cannot independently inspect or run the model as promised. The launch establishes Reflection's technical proposition and its intended market; the release will show what outsiders can actually build with it.
A large model, with a narrower cost claim
Reflection describes Beam as a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. The company says it trained the model on 23.8 trillion tokens and gave it a one-million-token context window. Those are company-reported specifications, not independent measures of usefulness in production.
The sharper commercial claim is efficiency. Laskin told Semafor that Beam needs three to four times less computing power than comparable open models to reason through a problem. Reflection also says Beam performs comparably to Z.ai's GLM-5.2 on advanced reasoning benchmarks. The claim is consequential if it holds: inference cost shapes how often a customer can use a model and whether running it on private infrastructure is affordable. But benchmark scores and a compute comparison do not by themselves establish end-to-end serving costs or performance across customers' workloads.
RuntimeWire's benchmark review found Beam near or ahead of some open models on selected coding tests and behind others. It also found that Reflection's compute comparison excludes several costs of serving a model. That scorecard makes the distinction between a model's benchmark showing and its operating economics particularly relevant to Reflection's pitch.
The founder's thesis meets the buying case
Laskin is selling control as much as capability. A company or government running an open-weight model can adapt it to internal data and infrastructure instead of depending entirely on a closed provider's service. Reflection's broader plan is to supply an "AI factory" for customers that want to train and deploy customized systems in private-cloud, on-premises or restricted environments.
That position serves several interests at once. Customers get another potential supplier; Reflection gets a route into organizations for which control and deployment location are procurement requirements; and its backers have exposure to demand for both AI models and the computing infrastructure they require. Nvidia, Sequoia Capital and Lightspeed Venture Partners are among Reflection's disclosed investors. In June, Reflection confirmed that its latest financing valued the company at $25 billion pre-money, according to Semafor.
The capital scale raises the bar for Beam. A model that is merely credible in a benchmark is not enough to turn a large valuation into a durable business. Reflection needs the model to perform well, the promised weights and documentation to make it usable, and its deployment offer to solve problems customers will pay to address. The company's reported funding and the investor names demonstrate that it has secured substantial backing; they do not demonstrate customer adoption or revenue.
The competition also runs in both directions. Reflection is positioning Beam against Chinese open models such as GLM and Alibaba's Qwen, while Western open-weight projects from companies including Thinking Machines Lab and Mistral are pursuing the same broad market. For Laskin, the point is to give customers an alternative they can own and customize. Whether Beam becomes that alternative depends on the model and tools developers can test after the promised weights arrive.
Beam is therefore a milestone for Laskin's founding thesis, not its proof. Reflection has moved from an argument about open AI and institutional control to a named model with published specifications and benchmark claims. The next test is whether people outside the company can use it, reproduce its performance and make the cost case work on real deployments.