Reflection AI Launches Beam, an Open‑Weight Model Targeting Chinese Rivals at Lower Compute Cost

Reflection AI, a two‑year‑old startup backed by Nvidia, Sequoia and Lightspeed, has officially released Beam, its first frontier open‑weight model. The company positions Beam as a “workhorse” for enterprises, public‑sector agencies and developers who need strong reasoning, coding and agentic capabilities without the heavy inference budget typical of today’s largest models.

Key specifications

• Total parameters: 501 B (23 B active)
• Pre‑training corpus: 23.8 trillion tokens
• Context window: 1 million tokens
• Architecture: text‑only mixture‑of‑experts with high‑compute reinforcement learning

Reflection claims Beam matches the advanced‑reasoning scores of Z.ai’s GLM‑5.2 and outperforms current Western open models while consuming roughly 3‑4× less inference compute. Independent verification is pending, but the reported efficiency could make Beam attractive for cost‑sensitive deployments.

Why it matters for GPU infrastructure

Beam’s lower compute footprint directly reduces the number of GPUs required per inference job, easing pressure on cluster capacity and cutting operational expenditure. For AI‑factory concepts championed by Nvidia, models like Beam enable sovereign or enterprise‑owned “AI factories” to run high‑quality workloads on smaller, more affordable GPU fleets — aligning with the push for open, locally‑controlled AI infrastructure.