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 launched Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, pretrained on 23.8 trillion tokens with a 1 million token context window, and plans to release weights and full technical details this month.
- Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and approaches Qwen3.8-Max on agentic tasks while using 3-4x less inference compute, though the performance claims haven't been independently verified.
- Reflection has raised roughly $4.7 billion from Nvidia, Sequoia Capital, and Lightspeed at a $25 billion pre-money valuation, and this summer signed deals worth more than $7 billion with SpaceX and Nebius to lock in Nvidia GB300 chip access through 2029.
- Beam outscores Mira Murati's Inkling on four coding benchmarks where both report results, though Inkling is multimodal and Beam is text-only, positioning Reflection against both Chinese open-weight models and Western closed labs like Anthropic and OpenAI.
- Reflection is pitching Beam and future models as the foundation for 'AI factories' — customized local AI systems for enterprises and sovereign nations — and is already testing the concept with South Korea's Shinsegae Group.
Why it matters: If Reflection's compute-efficiency claims hold up, the $25 billion-valued startup offers enterprises a Western open-weight alternative that undercuts the cost economics of Chinese models like GLM-5.2, while its $7 billion in locked-in Nvidia GB300 compute through 2029 directly targets the bottleneck that has kept Western open-weight efforts behind DeepSeek and Qwen.
Ask SkimNews
