Nvidia-backed Reflection AI challenges Chinese dominance in open-weight models

An Nvidia-backed US start-up has unveiled an open-weight model to challenge China’s dominance in freely available artificial intelligence, releasing a system that early third-party testing suggests could be one of the most token-efficient open models in its class.

Reflection AI on Monday debuted Beam, its first open-weight model built for coding, reasoning and agentic tasks.

The US firm said Beam was “competitive” with GLM-5.2 from Z.ai – also known as Zhipu AI – and was “approaching” Alibaba Group Holding’s Qwen3.8-Max on coding and agentic workflows, although Moonshot AI’s Kimi K3 remained ahead in “raw capability”. Alibaba owns the South China Morning Post.

Built on a “mixture of experts” approach, Beam distributes workloads among specialised networks. Reflection said this architecture allowed the model to activate just 23 billion parameters per task out of a total of 501 billion, substantially undercutting the compute demands of Z.ai’s GLM-5.2, which activated 40 billion out of roughly 744 billion parameters.

According to Reflection, Beam requires three to four times less inference compute than comparable open models to reason through problems, enabling users to run AI workloads faster and cheaper.

While official independent benchmarks have yet to be released, benchmarking platform Artificial Analysis said evaluation was under way. The firm noted that “early indicators suggest Beam will be one of the most token-efficient open models we’ve seen for its level of intelligence”.

The launch comes amid growing efforts in the US to counter the global rise of Chinese open-source AI developers, including DeepSeek, Alibaba, Moonshot AI and Z.ai.

In July, Nvidia and US tech giants including Microsoft, Palantir and Meta Platforms signed a public letter calling for the US to support a domestic open-weight AI ecosystem and avoid “premature restrictions” on downloadable models. The initiative also drew backing from IBM, OpenAI and Google.

Washington is also supporting the open-weight push. At a congressional hearing last month, US Treasury Secretary Scott Bessent urged American companies to develop more open-weight models, warning against allowing a handful of closed-model developers to dominate the industry.

Reflection has become part of that effort. The company has partnered with the Pentagon and US Department of Energy and has struck agreements to develop AI systems for US allies including South Korea.

Chinese open-weight AI developers have become increasingly embedded in the global ecosystem, with even US tech giants turning to their models to meet customer demand.

On Monday, Amazon said its cloud unit Amazon Web Services (AWS) added Z.ai’s GLM-5.3 to its Amazon Bedrock platform. AWS will share revenue with Z.ai based on model usage, according to Chinese tech media outlet Star Market Daily.

Reflection, founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, has raised over US$4 billion since its launch in 2024. It reached a US$25 billion valuation before releasing its first model.

Nvidia backed the start-up with US$800 million and provided it with vital chip access, alongside venture firms including Sequoia Capital, Disruptive and 1789 Capital, where Donald Trump Jnr is a partner.

Laskin said China’s grip on open-weight AI has left businesses and governments with limited Western alternatives for models they can actually control.

“The only way to own intelligence is, by definition, if it’s open,” he said in an episode of the Sources Podcast published Monday.

Beam is undergoing final red-teaming – where developers simulate cyberattacks and prompt exploits to test safety limits – and is available to a limited group of early users. Reflection plans to release the model weights later this month, alongside fine-tuning tools.