Reflection AI released Beam, a 501 billion parameter open-weight model, saying it matches the reasoning scores of China's GLM 5.2 while using three to four times less inference compute, according to the company's blog post. The Brooklyn, New York startup said it will release the model's weights under an Apache 2.0 license later this month.
Beam is a sparse mixture of experts model with 23 billion active parameters, pretrained on 23.8 trillion tokens with a context window extended to 1 million tokens, Reflection said. The company described Beam as "competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks," while acknowledging that "frontier open models like Kimi K3 remain ahead on raw capability."
Training involved more than 100 million reinforcement learning rollouts across 10,500 Nvidia GB300 GPUs over four weeks, drawing on 1.3 billion sandboxes and 1 million coding, agentic and STEM environments, Reflection said. Pretraining ran on 6,144 GB300 NVL72 GPUs in under four weeks.
Reflection has raised about $4.7 billion from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, reaching a $25 billion pre-money valuation, and has lined up more than $7 billion in compute deals with SpaceX and Nebius for access to Nvidia GB300 chips through 2029, TechCrunch reported. The company's performance claims have not been independently verified, TechCrunch noted.
An open-weight model that claims to match a Chinese frontier lab on reasoning while needing a fraction of the inference compute is a direct pitch to builders now paying per token for a closed model, or picking between GLM, Qwen and Kimi because nothing comparable ships with a Western license. Independent benchmarks, not Reflection's own numbers, will decide if that compute claim holds up.