Reflection AI on Monday, Oct. 5, unveiled Beam, its first open-weight language model, in a company blog post. Beam is a text-only mixture-of-experts model with 501 billion total parameters, 23 billion of them active. It was pretrained on 23.8 trillion tokens and has a context window of 1 million tokens, according to the company and TechCrunch (opens in new tab).
Reflection says that on advanced reasoning benchmarks Beam “achieves scores comparable to GLM-5.2 while using 3–4× less inference compute,” and that it outperforms leading Western open models. The company also says frontier open models such as Moonshot’s Kimi K3 “remain ahead on raw capability.” TechCrunch (opens in new tab) noted that Reflection’s performance claims have not been independently verified. Fortune (opens in new tab) pointed out that Z.ai released GLM-5.2 in July and followed it with GLM-5.3 in August. Reflection says its reinforcement-learning run generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks.
The model is not yet broadly available. “We will release the weights, technical report, model card, and developer artifacts later this month,” Reflection wrote, adding that the weights will carry an Apache 2.0 license. For now, a select group of users is getting early access through a waitlist while the model undergoes final red-teaming and evaluations.
Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin, its CEO, and Ioannis Antonoglou, its president and chief technology officer. It has raised about $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook data cited by TechCrunch (opens in new tab), and its last round valued it at $25 billion before the new money. Citing CNBC and Bloomberg, Fortune (opens in new tab) reported compute deals of $6.3 billion with SpaceX in June and $1 billion with cloud provider Nebius in July.
The company is pitching Beam to enterprises, governments and developers as a “workhorse” model and as the basis for what it calls “AI factories,” customized local systems trained on an institution’s own data. Its most direct U.S. rival may be Inkling, the open model that Mira Murati’s Thinking Machines Lab released in July, TechCrunch (opens in new tab) reported. Reflection’s own benchmarks show Beam ahead of Inkling on four coding tests where both report results, though Inkling is multimodal and Beam is text-only.