Tufra Labs Work with us

Get the most out of your GPUs.

Tufra Labs runs the optimization on your workload and delivers a verified patch, with the same models and the same outputs.

Work with us See results
Stock vLLM: 1,224 tok/sWith Tufra Labs: 2,444 tok/s Qwen3.8-27B on one H100, every user above 50 tok/s

01 · Results

Measured on real workloads.

Same model and GPU before and after, with the model’s outputs unchanged.

2,444tok/s

Throughput on one H100

Up from 1,224 with stock vLLM: 44 users instead of 22, each above 50 tokens per second.

$0.34 per million output tokens, half the stock cost.

One H100 SXM · vLLM 0.31, FP8, against stock defaults · reasoning on, chat and search prompts up to 12k tokens · every user above 50 tok/s · cost at $3 per GPU-⁠hour · outputs checked equivalent to stock

02 · How it works

LEVI evolves your stack and keeps only what’s faster.

  1. 01

    Runs on your cluster

    A few hours on your own GPUs. No new models, no new hardware.

  2. 02

    LEVI evolves your stack

    It proposes changes, learns from the ones that work and builds on them, from serving settings down to GPU kernels.

  3. 03

    Keeps the fastest

    Every change is verified to give the same outputs, and measured on your hardware.

03 · Use cases

Faster inference, shorter training, fuller GPUs.

LLM inference

Faster generation for chat, search and agents.

AI training

Shorter reinforcement learning and fine-tuning runs.

GPU utilization

More work from the GPUs you already own.

04 · Research

Built on our own research.

LEVI is cited in research from

Stanford University NVIDIA Meta UC Berkeley MIT Carnegie Mellon University

05 · Working with us

From first message to a faster stack.

You only pay if there’s a measured improvement.

  1. 01

    Tell us what you run

    Model, serving stack and GPUs.

  2. 02

    We run LEVI on your cluster

    Nothing leaves your machines.

  3. 03

    You keep the faster stack

    A verified patch, same outputs.

06 · Questions

What teams ask first.

What can you optimize?

The whole stack between your model and the GPU: serving settings, batching and scheduling, and GPU kernels, for inference and training. The results above are on vLLM and H100s; tell us what you run.

What access do you need?

Your model, a sample of real prompts, and access to a GPU machine like the ones you serve on. LEVI runs there.

Does our code or data go to an outside AI service?

No. LEVI runs its search with open-weight models on your own GPUs, so your code, prompts and data never leave your network.

How do you make sure nothing breaks?

Every change is checked against your current setup: the same outputs on your prompts, and measured on your hardware. You get the changes as a patch to review, and rolling back means reverting it.