Nvidia / 5 October 2026

Reflection Unveils Beam After Large Nvidia Training Run

Nvidia-backed Reflection AI has unveiled Beam, its first major model for coding, reasoning and agent workloads, with downloadable weights planned later this month after final evaluation. The company says one reinforcement-learning phase used 10,500 Nvidia GB300 GPUs for four weeks, providing a substantial disclosed example of Blackwell use by an open-model developer. Beam has 501 billion total parameters and activates 23 billion for each token. Reflection’s central claim is competitive performance with lower inference computation on selected comparisons, rather than leadership on every task. For Nvidia, the launch combines investment exposure with an operating hardware reference and the possibility of further inference demand wherever customers eventually choose to run the model.

Beam’s architecture is intended to concentrate computation on a subset of its network while retaining a much larger pool of learned parameters. Reuters places it among US efforts to compete with lower-cost Chinese open models, including those developed by Z.ai and Alibaba. Reflection’s published results show a mixed competitive position, with stronger models remaining ahead on several tasks. An open-weight release would let developers inspect, adapt and operate the model themselves, subject to the eventual licence and release terms. That differs from accessing a proprietary model exclusively through its developer’s service. It also makes the model’s practical serving requirements important, because users choosing control over deployment must supply the infrastructure and expertise to exercise it.

Reflection’s efficiency comparison needs a narrower interpretation than a claim that customers’ bills will fall by the same multiple. The company estimates generation computation using active parameters and generated tokens, and explicitly excludes some other work, including prompt processing and serving overhead. It reports roughly three to four times less inference computation than a comparison model on selected reasoning evaluations. That is a useful statement about a measured or estimated component of the workload, rather than a complete cost accounting. The eventual cost per successful coding task also depends on retries, tools, memory, utilisation and whether the model produces a correct result without extensive human repair or repeated attempts.

The commercial ambition is to give organisations more control over the intelligence they use. Reflection chief executive Misha Laskin told Axios that customers want to move “from renting it to owning it yourself.” An organisation can customise an open model around its own workflows, but taking control of the model does not eliminate the cost of computing or the need to maintain it. Reflection says it is already training its next model, indicating that Beam is part of a continuing development programme. For customers, that raises a familiar platform decision: use the hosted service for convenience, operate the released model for greater control, or combine the two as requirements change.

The Nvidia relationship is direct but should not be overstated. Reflection is backed by the chipmaker and has disclosed large-scale use of its GB300 hardware; those facts establish an investment link and an actual computing workload. They do not reveal how much Nvidia earned from this training run or the share of future Beam deployments that will use its systems. The planned weight release could expand the number of organisations able to host the model, while the company’s own future training requires further capacity. Both channels matter to Nvidia, but their value rests on practical adoption. A model that is efficient on selected tests still has to earn a place in customer applications that run repeatedly and at meaningful scale.