Reflection AI ships Beam, its first public model, after two years and $25B
The Nvidia- and Sequoia-backed open-weight lab finally has something to show: a model it says beats other Western open models at coding and agentic tasks on 3-4x less compute, built explicitly as a non-Chinese option for sovereign AI buyers.
Reflection AI released its first public model, Beam, on October 5 -- the open-weight lab's first shipped product since founders Misha Laskin and Ioannis Antonoglou (both ex-DeepMind) started the company in 2024. Reflection says Beam needs 3-4x less compute than comparable open models to do reasoning tasks, and is particularly strong at coding and agentic work. On Reflection's own benchmarking, Beam lands roughly at the level of Z.ai's GLM-5.2 (released June 2026) and approaches Alibaba's Qwen 3.8-Max -- ahead of other Western open models, but still behind OpenAI and Anthropic's top closed models.
A $25 billion valuation with nothing shipped, until now
Reflection raised at a $25 billion pre-money valuation in June 2026 with Nvidia, Sequoia, and Citigroup among its backers, on the strength of a pitch rather than a product -- the company had not released a model or a weights checkpoint before Beam. CEO Misha Laskin frames the gap Reflection is trying to fill as a sovereignty problem for enterprises and governments that want a capable open model without depending on Chinese labs like DeepSeek or Alibaba: "They don't really have very good options today." On AI governance, Laskin's position is pointedly not anti-regulation: "You want multiple voices around the table, both open and closed."
Why a late, mid-tier model still matters for buyers
Beam isn't a frontier model, and Reflection says so implicitly by benchmarking it against open, not closed, competitors. What makes it worth tracking is the category it's trying to create: a US-aligned open-weight option for organizations whose procurement rules or sovereignty requirements rule out Chinese open models but whose budgets or deployment constraints rule out the closed frontier. Whether "3-4x less compute" holds up under independent benchmarking, and whether that's enough to justify two years and $25 billion against a company that shipped nothing until now, is exactly the kind of claim worth checking against real usage rather than Reflection's own numbers once enterprises actually start running it.