Reflection

Artificial Intelligence · Open-Weight Models · Developer Tools

Launch

Reflection launches Beam, a 501B-parameter open-weight model

Activating only 23 billion of its parameters per token, it aims to match larger Chinese rivals at a fraction of the compute cost.

• Reflection AI unveiled Beam on October 5, 2026, its first open-weight model: a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic tasks.
• Reflection AI, founded in 2024 by former Google DeepMind researchers, has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, at a $25 billion pre-money valuation.
• Beam was pretrained on 23.8 trillion tokens, then put through a high-compute reinforcement learning run that generated over 100 million rollouts across 10,500 Nvidia GB300 GPUs over four weeks.
• Reflection says Beam scores comparably to Z.ai's larger GLM-5.2 (744B total, 40B active parameters) on advanced reasoning benchmarks while using three to four times less inference compute, though no independent evaluation exists yet.
• Open-weight model leaderboards have recently been dominated by Chinese labs like DeepSeek, Qwen, GLM, and Kimi; Reflection explicitly frames Beam as advancing a 'Western open-weight frontier.'
• Reflection plans to release Beam's weights, technical report, and model card under an Apache 2.0 license later in October 2026, with distribution through hyperscalers and neoclouds.
• Beam is a bid by a heavily capitalized US startup to prove that inference efficiency, not sheer scale, can keep Western labs competitive with China's fast-moving open-weight AI ecosystem.