Dynamic Beads Introduces Nested Run to Lower Cost of Running Large AI Models
News related to:Dynamic Beads · 2 min read
Dynamic Beads, a company that builds and licenses plug-and-play upgrades for AI models, has introduced Nested Run, a software upgrade designed to make large AI models run faster and at a lower cost. The new upgrade, which is currently compatible with the Qwen3.8-27B model running on NVIDIA H100 and H200 GPUs, offers significant performance improvements and cost savings.
According to the company, Nested Run can read long prompts up to 1.46 times faster than the original model. Additionally, it can process prompts of 16,000 tokens 1.3 times faster within vLLM, a widely used open-source serving engine. The upgrade also allows follow-up questions about the same document to be answered at a cost up to 74 percent less when the saved document state is reused.
For larger inputs, such as prompts of 128,000 tokens, the upgrade enables them to fit on a single 80 GB GPU and run 1.51 times faster. The model's embedding table, a large block of GPU memory, is reduced by 73 percent, making it more efficient. The company notes that the upgrade can return the original model's answers word for word when needed, ensuring consistency in responses.
Eytan Goldstein, the inventor of Nested Run, explained that the approach is designed to give each word in a prompt the effort it actually needs, allowing the same model to run faster without changing what it knows. The upgrade installs alongside a company's existing model and server software without requiring retraining of the model from scratch.
Dynamic Beads also offers Nested Run+, which includes additional tools to speed up the entire request, including answer writing. These tools can make complete requests up to 2.2 times faster, reduce model memory by about 45 percent, and cut reasoning length by 18 percent.
The company estimates that the upgrade can save about 10 percent on mixed traffic and about 25 percent on long-document traffic.
Dynamic Beads is inviting GPU cloud providers and AI platforms to measure the savings on their own traffic before committing.
The company's first pack supports the Qwen3.8-27B model, and packs for additional model families are in development. A patent application covering the method has been filed with the United States Patent and Trademark Office.