Beam is Reflection AI's first open-weight model. On October 5, 2026, the two-year-old U.S. startup officially unveiled Beam: 501 billion total parameters with only 23 billion active per inference, pretrained on 23.8 trillion tokens, with a 1-million-token context window, positioned squarely against Chinese open models such as Z.ai's GLM-5.2, DeepSeek, and Qwen. The weights and full technical details are due later this month, distributed through cloud providers and open-source libraries.
The pitch is the per-token bill, not the parameter count
Beam is a text-only mixture-of-experts model trained with heavy reinforcement learning for reasoning, coding, and agentic tasks. In Reflection's own testing, it scores roughly on par with GLM-5.2 on advanced reasoning benchmarks, ahead of today's leading Western open models, while using a third to a quarter of the inference compute. The parameter comparison shows the bet: GLM-5.2 carries about 744 billion total parameters with roughly 40 billion active, and Beam chases similar scores with far fewer active parameters, wagering that enterprises pay for cost per completed task, not spec sheets. All of these figures are vendor-reported and not yet independently verified; the real gap will only be clear once the weights ship and the community runs its own evaluations.
Why a U.S. startup plans to make money by giving weights away
Reflection was founded by two former Google DeepMind researchers and counts Nvidia, Sequoia, and Lightspeed among its backers, with roughly $4.7 billion raised per PitchBook and a recent valuation around $25 billion. This summer it signed compute deals worth more than $7 billion in total, locking up Nvidia GB300 capacity from SpaceX and Nebius through 2029. Its business model differs from selling API access: it wants to build localized "AI factories" for enterprises and sovereign nations, where customers train their own systems on proprietary data using open weights — a sovereign AI partnership with South Korea's Shinsegae Group is already being tested. That model only works if a capable open model gathers the ecosystem first, which is exactly Beam's job. The Kolibri open-weight release we covered earlier follows the same logic: open weights, paid deployment and customization.
What can actually be concluded today
Three things are confirmed: the Western camp finally has a model whose scale and positioning directly challenge China's top open tier; if independent tests confirm the efficiency story, per-task inference prices at this level will come under pressure; and Beam is text-only for now, one modality behind some Western rivals. What is not confirmed is the benchmark sheet. Until weights and evaluation details are public, the prudent move for teams evaluating models is to put Beam on the watch list and run their existing eval sets when it lands, rather than re-architecting around it in advance.
