Patent-pending software measured up to 1.46 times faster on long prompts and 1.3 times faster in vLLM, with a free first month for AI providers
NEW YORK, NY, UNITED STATES, October 30, 2026 /EINPresswire.com/ — Dynamic Beads, an emerging artificial intelligence (AI) infrastructure company, today introduced Nested Run, a plug-and-play software upgrade designed to make large language models (LLMs) faster and less expensive to run on the graphics processing units (GPUs) that companies already operate.
In tests on the open-weight Qwen3.8-27B model running on NVIDIA H100 and H200 GPUs, comparing the original model with the upgraded model on the same hardware, Dynamic Beads measured the following results:
– Long prompts were read up to 1.46 times faster.
– Inside vLLM, a widely used open-source serving engine, prompts of 16,000 tokens were ready 1.3 times faster.
– Follow-up questions about the same document cost up to 74 percent less when the saved document state is reused.
– Prompts of 128,000 tokens fit on a single 80 GB GPU and ran 1.51 times faster.
– The model’s embedding table, a large block of GPU memory, was 73 percent smaller.
An exact mode returns the original model’s answers word for word whenever a customer requires it. In a head-to-head test at the same amount of work, giving each word its own depth kept about 10 points more agreement with the original model than cutting every word at the same layer, the approach used by methods such as SwiftKV, including 25 points more on conversational text.
The upgrade installs alongside a company’s existing model and server software, without retraining the model from scratch. The company says the savings matter most for long inputs such as documents, search results, coding assistants and extended conversations, where reading the prompt makes up a large share of GPU time. Based on its measured results, Dynamic Beads estimates savings of about 10 percent on mixed traffic and about 25 percent on long-document traffic.
“Every company running AI models pays for GPU time on every request, and those bills keep growing,” said Benny Goldstein, Founder of Dynamic Beads. “Our goal is simple: the same model, the same answers, and a lower bill. We believe Dynamic Beads can become one of the companies that defines how AI is run efficiently in the years ahead.”
The approach was invented by Eytan Goldstein. It is designed to plug into existing open-weight AI models. The first pack supports Qwen3.8-27B, 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.
“The idea is that not every word in a prompt needs the same amount of work,” said Eytan Goldstein, Inventor of Nested Run. “Giving each word the effort it actually needs is what lets the same model run faster without changing what it knows.”
Dynamic Beads also offers Nested Run +, which adds optional tools that speed up the whole request, including answer writing. In company tests these tools made complete requests up to 2.2 times faster, reduced model memory by about 45 percent and cut reasoning length by 18 percent.
The company is inviting GPU cloud providers and AI platforms to measure the savings on their own traffic before committing. “We offer a free first month at no cost,” said Steve Marks, Chief Partnerships Officer of Dynamic Beads. “We run the upgrade side by side with a provider’s current setup, on its real requests, and hand over a written savings report. The numbers speak for themselves.”
Nested Run and Nested Run + are licensed to the companies that run AI models, either as a license for a provider’s fleet or as a share of the measured savings. More information is available at https://dynamicbeads.com.
About Dynamic Beads
Dynamic Beads builds and licenses plug-and-play upgrades that make AI models faster and less expensive to run on existing GPUs. Its first products, Nested Run and Nested Run +, are designed for GPU cloud providers, AI inference platforms and companies that serve open-weight large language models. The company’s method is patent pending. Dynamic Beads was founded by Benny Goldstein, with inventor Eytan Goldstein and Chief Partnerships Officer Steve Marks.
Benny Goldstein
Dynamic Beads
bennygoldstein@msn.com
Legal Disclaimer:
EIN Presswire provides this news content “as is” without warranty of any kind. We do not accept any responsibility or liability
for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this
article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
Media gallery
