Reflection AI just raised the stakes in one of the more interesting corners of the AI race: open-weight models built in the West that can genuinely compete with Chinese labs. According to TechCrunch, the company has officially launched Beam, its first frontier open-weight model, claiming it matches leading Chinese open models on advanced reasoning benchmarks while using three to four times less inference compute. That’s a bold claim. It hasn’t been independently verified. But if it holds up, it matters a lot.
Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and has a 1 million token context window. For reference, Z.ai’s GLM-5.2 runs around 744 billion total parameters with 40 billion active. Reflection says Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and outperforms leading Western open models, all at significantly lower inference cost. The company describes it as a “workhorse model” aimed at enterprises, developers, and the public sector, trained with high-compute reinforcement learning for reasoning, coding, and agentic tasks.
The competitive picture here is crowded. Reflection is positioning Beam against closed labs like Anthropic and OpenAI, Chinese open models like DeepSeek and Qwen, and Western open-weight players including Mistral, Meta, and Cohere. Its closest U.S. comparison is probably Inkling, the open model from Mira Murati’s Thinking Machines Lab that launched in July. Reflection’s own benchmarks show Beam outscoring Inkling on four coding tests, but Inkling is multimodal and Beam is text-only, so they’re not a clean apples-to-apples comparison.
What makes Reflection worth watching isn’t just the model. It’s the infrastructure play behind it. The company was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with a reported pre-money valuation of $25 billion. This summer, Reflection signed deals worth over $7 billion with SpaceX and Nebius to lock up access to Nvidia’s GB300 chips through 2029. That’s a serious bet on needing raw compute at scale for a long time.
The go-to-market strategy is also notable. Reflection is pitching what it calls “AI factories” to enterprises and sovereign nations, essentially letting institutions train Beam on their own proprietary data to build customized, locally deployed AI systems. Nvidia CEO Jensen Huang has pushed this idea for years, and given that Nvidia is a backer, there’s an obvious alignment of interests. Beam’s first real sovereign test is already underway: Reflection has begun a pilot partnership with Shinsegae Group in South Korea. Axios also reported that hedge funds and trading firms are among those interested in building similar systems.
The model’s weights and full technical details are expected to be released later this month, with distribution planned through hyperscalers and neoclouds, plus integrations across open source libraries. That release timeline is the real test. Performance claims on self-reported benchmarks are one thing. Community scrutiny of actual weights is another. Still, the combination of serious funding, locked-up compute, and a clearly defined enterprise and sovereign buyer strategy puts Reflection in a different category than most AI startups making noise at launch. The next few weeks will show whether Beam can back it up.
