
Reflection AI has moved from promise to product. A day after reports that the Nvidia-backed company was preparing a US open-weight contender, it introduced Beam, a model built for coding, reasoning and AI agents. It has 501 billion parameters in total, although only 23 billion are active for each token because it uses a mixture-of-experts design.
That distinction matters more than the enormous headline number. A mixture-of-experts model routes work to part of its network rather than using every parameter for every answer, potentially making inference less expensive. Reflection says Beam is designed to deliver strong results with less compute than some much larger open-weight rivals. For businesses considering self-hosted AI, the eventual serving cost may matter as much as a benchmark score.
There is an important catch. Beam has been announced, but its weights are not yet available for general download. Reflection says it is finishing red-teaming and evaluation, is offering access to a select early group and plans to release the weights, technical report, model card and developer tools later in October. It says the weights will carry an Apache 2.0 licence. Until those materials arrive, outside developers cannot fully inspect the release or reproduce the company’s performance claims.
Reflection’s own results put Beam in contention on several coding and agentic tests, but they do not show a clean victory over the strongest Chinese models. The company acknowledges that Kimi K3 is still ahead on raw capability and argues instead that Beam’s advantage is efficiency at inference time. Those comparisons are company-reported, and real-world performance will need independent testing across different tasks, languages and hardware setups.
The scale of the project is striking. Reflection says it pretrained Beam on 23.8 trillion tokens and ran reinforcement learning using roughly 10,500 Nvidia GB300 GPUs for four weeks, generating more than 100 million rollouts. Those numbers underline why a model marketed as open weight can still require resources far beyond the reach of most startups to create. Opening the finished weights is a different proposition from opening the training pipeline.
For enterprise customers, the practical appeal is choice. A downloadable model could let a bank, hospital or government agency keep sensitive workloads on infrastructure it controls and adapt the system to local needs. But the promised Apache licence, documentation, safety material and deployment tools will determine how useful Beam actually is. Our guide to open-weight AI explains why access to model weights does not automatically mean complete openness.
This is also a strategic moment for the US AI sector. Chinese open-weight models have given developers capable and often cheaper alternatives to closed American services. Beam is Reflection’s attempt to offer another option without requiring every customer to depend on a proprietary API. The announcement makes that contest more concrete, but the meaningful test starts when people outside the company can run the model, inspect the licence and compare its actual cost with the alternatives.







