Reflection AI Unveils Open-Weight Beam Model To Rival China

Reflection AI, a two-year-old startup backed by Nvidia, unveiled Beam on October 5, an open-weight artificial intelligence model the company says can match leading Chinese open systems on advanced reasoning and coding benchmarks while using far less computing power. The New York-based firm said it would release Beam's weights under an Apache 2.0 licence later this month.¹²³

A model built for efficiency

Reflection AI describes Beam as a Mixture-of-Experts system that activates only 23 billion of its 501 billion total parameters at a time, designed to handle coding, reasoning and agentic tasks. The company said training relied on 23.8 trillion curated tokens run across 6,144 Nvidia GB300 NVL72 processors, finishing in less than a month and hitting a utilization rate of 92.3 percent "goodput" by the end.²

Reflection AI said the model's reinforcement-learning phase generated more than 100 million rollouts on 10,500 GB300 GPUs over four weeks, sustaining an average of 110,000 concurrent rollouts, which the company described as "one of the largest scale RL runs conducted by any open lab to date". SiliconANGLE reported that, despite 71 errors logged during training, the system typically bounced back within eight minutes, and that Reflection AI had constructed almost all of its infrastructure internally.²⁶

According to TechCrunch, Beam equals top Chinese open models' results on difficult reasoning tests while needing three to four times less computing power at inference, and it surpasses current leading Western open models, beating Thinking Machines Lab's Inkling on four comparable coding benchmarks.¹

Measured against Chinese rivals

Reflection AI said Beam performs on par with Z.ai's GLM-5.2 while requiring a fraction–three to four times less–of the inference compute, and is closing in on Alibaba's Qwen3.8-Max for coding and agent-based tasks, though Moonshot AI's Kimi K3 still leads on sheer capability. "Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time," the company said.²⁵

The South China Morning Post said Beam needs only a quarter to a third as much computing power as similar open models to work through problems, portraying the launch as a direct bid to counter China's hold on the open-weight segment. Fortune noted that China is currently dominating open-source AI competition, citing OpenRouter leaderboard rankings in which Nvidia's Nemotron model sits at No. 6 and OpenAI's gpt-oss at No. 19.⁵⁴

The Hill reported a claim, attributed to Reflection AI, that 80 percent of developers worldwide using open-source tools build with Chinese models, and cited Airbnb chief executive Brian Chesky as saying the company had been "relying a lot on Alibaba's Qwen model and described it as very good and fast and cheap".⁸

Funding and compute deals

The New York Times reported that Reflection AI, which has Nvidia's backing, was launched in March 2024 by Misha Laskin and Ioannis Antonoglou, two alumni of Google DeepMind now serving as chief executive and as president and chief technology officer respectively. TechCrunch reported the company has raised roughly $4.7 billion, with its most recent funding round valuing it at $25 billion pre-money.³¹⁴

TechCrunch reported that Reflection AI struck agreements exceeding $7 billion combined with SpaceX and Nebius, locking in use of Nvidia's GB300 chips until 2029. Fortune broke down the arrangements individually, noting a June contract worth $6.3 billion with SpaceX and a separate $1 billion July agreement with Nebius, in addition to a May partnership to provide models to the U.S. Department of Energy and the U.S. Department of War.¹⁴

Company's stated rationale

Reflection AI said it intends Beam to serve as a "workhorse model for enterprises, the public sector, and developers" and said it is "already training the next, larger model on our road map". The company argued that open systems "evolve faster, attract broader participation, and foster ecosystems of research and innovation that no single organization can achieve alone" and are "safer" because openness allows more researchers to identify vulnerabilities.¹⁴