Reflection AI, the two-year-old lab that raised billions of dollars before releasing a single model, has at last put one in the open. On October 5 the company introduced Beam, a 501-billion-parameter open-weight model aimed squarely at the Chinese systems that have come to define the open end of the market. The weights will be published under an Apache 2.0 licence later this month, according to reporting from Help Net Security.
Beam is a text-only mixture-of-experts model. Of its 501 billion parameters, roughly 23 billion fire for any given token, which is what keeps the cost of each answer down. Reflection says it pre-trained the model on 23.8 trillion tokens and gave it a one-million-token context window, then leaned on high-compute reinforcement learning to sharpen its reasoning, coding and agentic behaviour.
Capable, but not the fastest car on the track
The benchmark numbers are respectable rather than record-breaking. Beam scores 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench v2.1 and 97.8 on the AIME 2026 maths set. On most published rows it trails the current leaders among open models, including DeepSeek V4.1 Flash, Moonshot's Kimi K3 and Z.ai's GLM 5.3. Reflection does not hide this. Its pitch is about economics, not leaderboards.
The company claims Beam matches the older GLM-5.2 on advanced reasoning while using three to four times less compute at inference time, with a wider gap against the two-trillion-parameter class of models such as Alibaba's Qwen 3.8-Max. In plain terms, Reflection is arguing that a company can get most of the capability for a fraction of the serving bill, which matters a great deal to any business running these systems at scale.
A Western answer to the open-model question
Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. Nvidia led an $800 million slice of a 2025 round that valued the startup at around $8 billion, and later funding talks pushed the implied figure far higher. The lab has signed large compute commitments, including a multi-year arrangement to use Nvidia chips hosted at SpaceX's Colossus 2 facility.
From the start the company framed itself as an American, open-weight counterweight to DeepSeek and Qwen, the argument being that Western enterprises and governments want a capable base model they can inspect and host themselves without routing their work through Chinese systems. Beam is the first real test of that thesis. It arrives into a crowded field that already includes DeepSeek's open releases, smaller efficient entrants such as NaiveAI's n05 Flash, and translation-focused work like Bilibili's Index-Translate.
Whether efficiency alone is enough to win developers who can download a stronger model for free is the open question. What is no longer in doubt is that Reflection has something to ship. After years of funding headlines and no product, that is the news.
Commentarii · 0