# Dynamic Beads

> Dynamic Beads is an emerging leader in AI inference savings. Its patent-pending, plug-and-play packs make the large language models companies already serve faster and cheaper to run on their existing GPUs: up to 1.46x faster long-prompt reading, 74% cheaper follow-up questions, 128k-token prompts on a single 80 GB GPU, and 1.3x faster reading inside vLLM, with an exact mode that keeps the original answers. First release: Qwen3.8-27B.

Website: https://dynamicbeads.com/ · Contact: bennygoldstein@msn.com · Patent pending.

## Products
- **Nested Run** (core): up to 1.46x faster long-prompt reading (64k tokens); 1.3x faster reading inside vLLM on one H100 (1.28x at 4k, 1.30x at 16k, 1.25x at 64k); follow-up questions on the same document up to 74% cheaper (measured at 16k when the saved state is reused); 1.51x faster on 128k-token prompts on a single 80 GB GPU; 73% smaller embedding table; exact mode returns the original model's answers word for word.
- **Nested Run +**: everything in Nested Run plus add-ons: up to 2.2x faster whole requests, 1.3x faster answer writing, about 45% smaller model memory, 18% fewer reasoning tokens, up to 2.1x faster edits and copy-heavy answers, 81% cheaper repeated questions.

## Who it is for
GPU cloud providers, AI inference platforms and companies serving open-weight large language models, especially long-input traffic: documents, retrieval (RAG), coding agents and long conversations.

## How to buy
Free first month: side-by-side measurement on the customer's own traffic with a written savings report. Then a license for the customer's fleet, or a share of the measured savings. Runs on the customer's own GPUs (container, vLLM add-on) or as a hosted endpoint on request.

## Facts
- At the same amount of work, giving each word its own depth beat cutting every word at the same layer (the SwiftKV-style approach) by about 10 points of agreement with the original model on average, 25 points on chat and 22 points on sums. (Teacher-forced next-word agreement, Qwen3.8-27B.)
- First supported model: Qwen3.8-27B. Similar savings are projected for other models; more packs are coming.
- Measured on NVIDIA H100 and H200 GPUs, original vs upgraded model, same hardware and software.
- Team: Benny Goldstein (Founder), Eytan Goldstein (Inventor), Steve Marks (Chief Partnerships Officer).