A cost-focused frontier model
Reflection AI, a two-year-old Brooklyn startup founded by ex-Google DeepMind researchers, has formally unveiled Beam, its first frontier open-weight model, TechCrunch reports. The pitch is not simply raw capability; it is efficiency. Beam is a text-only mixture-of-experts system with 501 billion total parameters and 23 billion active parameters, pretrained on 23.8 trillion tokens and offering a 1 million token context window. For comparison, Z.ai's GLM-5.2 carries roughly 744 billion total and 40 billion active parameters.
The startup says it tuned Beam with high-compute reinforcement learning for reasoning, coding, and agentic work. Reflection's own benchmarks claim Beam scores on par with GLM-5.2 on advanced reasoning while using 3-4x less inference compute. Those figures are self-reported and not yet independently verified, but if they hold, Beam could reshape cost expectations for deploying frontier-scale reasoning.
Positioning against Chinese and Western rivals
Reflection is deliberately drawing a line: a Western open-weight model that can compete with Chinese labs without matching their compute budgets. The company positions Beam against:
- Chinese open models from DeepSeek, Qwen, and Z.ai
- Western open players such as Mistral, Meta, and Cohere
- Closed labs like Anthropic and OpenAI
Its most direct U.S. comparison is Inkling, the open model from Mira Murati's Thinking Machines Lab. Reflection says Beam outscores Inkling on four coding benchmarks where both report results, though Inkling is multimodal and Beam is text-only. The startup has raised about $4.7 billion from Nvidia, Sequoia Capital, and Lightspeed Venture Partners at a $25 billion pre-money valuation, and it has locked in more than $7 billion in compute deals with SpaceX and Nebius for Nvidia GB300 chips through 2029.
The AI factory strategy and what remains unproven
Beyond selling model access, Reflection is pushing "AI factories" — local, customized systems built on an institution's own proprietary data. TechCrunch reports that enterprises, sovereign nations, hedge funds, and trading firms are prime targets, with an early partnership test underway with Shinsegae Group in South Korea. The company will release Beam's weights and full technical details this month, with distribution through hyperscalers and neoclouds at launch.
The biggest open question is verification. Reflection's benchmark claims have not been independently confirmed, and the field is crowded with aggressive self-reported numbers. But the cost angle is concrete: if Beam really delivers GLM-5.2-class reasoning at a fraction of inference compute, it offers a meaningful option for organizations that want frontier performance without sending data or money to Chinese labs. For now, TechCrunch notes Reflection did not respond to requests for additional information before publication.