Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Reflection AI unveiled Beam, a 501-billion-parameter open-weight mixture-of-experts model with 23 billion active parameters, 23.8 trillion training tokens, and a 1 million token context window, targeting reasoning, coding, and agentic tasks.
- Beam is claimed to match Z.ai's GLM-5.2 on advanced reasoning benchmarks and outperform Western open models while using 3-4x less inference compute, though the performance claims have not been independently verified.
- Reflection positions Beam against closed labs (Anthropic, OpenAI), Chinese open developers (DeepSeek, Qwen, Z.ai), and Western open-model peers including Mistral, Meta, Cohere, and Mira Murati's Thinking Machines Lab, whose Inkling model Beam reportedly outscores on four coding tests.
- Reflection has raised roughly $4.7 billion from Nvidia, Sequoia, and Lightspeed at a $25 billion pre-money valuation, and this summer signed compute deals collectively exceeding $7 billion with SpaceX and Nebius to lock down Nvidia GB300 chip access through 2029.
- Reflection is pitching Beam to enterprises and sovereign nations as the foundation for customizable 'AI factories' trained on proprietary data, and is already testing the concept with South Korea's Shinsegae Group.
Why it matters: Reflection is stacking a $4.7B raise, a $7B+ compute lockup through 2029, and Nvidia's backing against a model it claims matches Chinese open-model performance at a fraction of the inference cost — if the benchmarks hold, enterprises and governments shopping for open-weight AI now have a Western alternative to Z.ai and Qwen that could undercut Mistral, Meta, and even Murati's Inkling on compute economics.
Ask SkimNews
