Now running inside vLLM · Patent pending

Run the same AI.
Pay less for every token.

1.46×
faster prompt reading on long inputs (64k tokens)
74%
lower cost for follow-up questions on the same document*
1.51×
faster on 128k-token prompts, on a single 80 GB GPU
73%
smaller embedding table in GPU memory

What one upgrade does to a 27B model. Original vs upgraded, same GPUs, same software. Measured on Qwen3.8-27B, NVIDIA H100 and H200.

Dynamic Beads builds and licenses plug-and-play upgrades for the AI models you already serve. Same model, same answers, measurably faster and cheaper on the GPUs you already have.

🎁 Get a free month at no cost. See the proof for yourself.

Estimate your savingsSee the products

* Follow-up saving measured at 16k tokens when the follow-up reuses the saved document state. Results are updated as new test rounds finish. An exact mode returns the original model's answers whenever you need them.

Similar savings are projected for other models; packs for more models will be released shortly.

Same request, side by side

A 16,000-token prompt in vLLM

Same model, same NVIDIA H100, same server software. Time until the first word of the answer.

Original model
0.00 s
With Nested Run
0.00 s
Measured: 1.28× at 4k · 1.30× at 16k · 1.25× at 64k · 1.32× with 4 requests at once
Savings calculator

What would you save each month?

Pick your fleet and your traffic. The biggest savings come from long inputs: documents, search and retrieval, coding agents and long conversations.

Modeled from measured test results at full-month utilisation. Your own savings are measured on your traffic before you pay anything.

Estimated savings per month
$21,170
$254,040per year
$211,700current monthly GPU bill
Products

Two packs. One install.

Each pack is built and tested for a specific model and licensed for your whole fleet. The first supported model is Qwen3.8-27B; more models follow.

NESTED RUN

Nested Run

The core upgrade. Your model reads faster and costs less, and an exact mode returns the original answers on demand.

  • Long-prompt readingup to 1.46× faster
  • Inside vLLM, same GPU1.3× faster
  • Follow-up questionsup to 74% cheaper
  • Longest prompts per GPU128k on 80 GB
  • Embedding memory73% smaller
  • Exact modeoriginal answers
NESTED RUN +

Nested Run+

Everything in Nested Run, plus the add-ons that speed up the whole request, answers included.

  • Whole request (prompt + answer)up to 2.2× faster
  • Answer writing with compact weights1.3× faster
  • Model memory~45% smaller
  • Shorter reasoning, more right answers18% fewer tokens
  • Edits and copy-heavy answersup to 2.1× faster
  • Repeated questions81% cheaper

More upgrades are in the lab and will be added here as their tests complete.

Runs where you run

Your servers or ours

ON YOUR INFRASTRUCTURE

Install it in your stack

A ready container and a vLLM add-on for your own GPUs, in your cloud or data center. Your data never leaves your servers.

HOSTED · ON REQUEST

Use our endpoint

Point your application at a hosted, OpenAI-compatible endpoint and start saving without touching your servers.

vLLMHugging Face TransformersNVIDIA H100NVIDIA H200Qwen3.8-27BMore models coming
Built for production

What your team will ask first

ANSWERS

Exact mode

Flip one switch and get the original model's answers, word for word.

PRIVACY

Your data stays put

Runs inside your environment. Nothing is sent to us.

EFFORT

No retraining

Your model is not retrained from scratch. A short calibration is included.

PROOF

Measured on your traffic

Paired tests on your own requests, with the numbers in a report you keep.

How it works for you

Proof first. Then you pay.

01

Free first month

We run the upgrade beside your current setup on your real traffic for a month, at no cost, and show you the difference.

02

Install

A ready container for your servers, or a hosted endpoint. No retraining your model from scratch.

03

Save

A monthly savings report. License it, or share the savings with us: you choose.

Questions

Frequently asked

Will my model give different answers?

Exact mode returns the original model's answers word for word. In fast mode, answers stay close to the original, and we measure that on your own traffic before you decide.

Which models are supported?

The first pack is for Qwen3.8-27B. Packs for more model families are in testing. Tell us what you serve and we'll tell you where it stands.

Does it work with vLLM?

Yes. In our tests inside vLLM on an H100, a 16,000-token prompt was ready 1.3× faster, and exact mode matched normal vLLM on every test prompt.

Where do the biggest savings come from?

Long inputs: documents, retrieval, coding agents and long conversations. Short chat saves less. The calculator above shows both.

How is it priced?

Either a license, or a share of the savings we measure. You choose after the pilot.

Is there a free trial?

Yes. Your first month is free, at no cost. We measure side by side on your traffic and give you a written savings report. See the proof for yourself.

About

The people behind Dynamic Beads

BG

Benny Goldstein

Founder

Benny leads Dynamic Beads. A lawyer by training (LLB, LLM) specialising in patents and contract law, with years of investor relations and capital raising.

EG

Eytan Goldstein

Inventor

Eytan invented Nested Run, the method behind every Dynamic Beads product. Patent pending.

SM

Steve Marks

Chief Partnerships Officer

Steve leads customer and business relations. He brings many years of experience building lasting client partnerships, and is part of the team that created Dynamic Beads.

For engineers

Technical notes

Per-word depth beats a uniform cut

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. Measured on Qwen3.8-27B with our own matched-work reimplementation of uniform layer skipping.

LLM inference optimizationPrefill accelerationTime to first token (TTFT) Per-token adaptive depthLong-context inference (128k)KV cache reuse Prefix cachingvLLM integrationGPU cost reductionTokens per GPU-hour RAG cost reductionEmbedding table compressionExact-match fallback mode Hybrid linear-attention modelsNVIDIA H100 / H200Qwen3.8-27B Open-weight LLM servingAI inference cost per million tokens
Contact

See your savings on your own traffic

Tell us what you serve. Get a free month at no cost and see the proof for yourself.

Prefer email? Write to us directly at bennygoldstein@msn.com