SOLISTIQ
Book a call

Technology / Q‑Foundry

Many products.
One shared model.

The world’s first foundation model for inventory management.

Q‑Foundry learns demand patterns across products and uses that knowledge for new products – without training a model for each one. Developed from research at the University of Würzburg.

The result

24% lower forecast error.

Forecast error · Baseline = 100
  1. Baseline
  2. Models trained per product −13%
  3. Q‑Foundry −24%

Q‑Foundry 2.5 · Pretrained on 1.2M products. Lower error is better.

In four steps

From learned patterns to a concrete order.

Four steps lead from the sales histories of many products to your order: learn, understand, forecast, plan. Here they are at a glance; further down, each step in detail.

  1. 01 Learn
  2. 02 Understand
  3. 03 Forecast
  4. 04 Plan

01 Pretraining

Prior knowledge from millions of products.

Q‑Foundry learns demand patterns across products. That knowledge is available for new products, too.

02 In use: your product

A forecast for your product.

Your sales history and available demand drivers determine how the model applies its prior knowledge.

03 In use: the forecast

Demand matched to your service target.

Quantiles describe the demand quantity to plan for at your chosen protection level.

04 In use: order planning

Orders under your conditions.

Optimization combines the forecast with inventory, lead times and ordering rules.

01Learn

One model, trained on millions of products.

With individual models, a separate model is trained for each product. Q‑Foundry learns across products: pretraining teaches it patterns from millions of examples. Each product still gets its own forecast.

2.5M
products
Sales histories from a range of industries.
1.5B
data points
Historical data used in pretraining.
1 model
Q‑Foundry
For all products, without client-specific training.
Before: individual models

Forecasts

Each model learns only from its own product’s history.

Today with Q‑Foundry

Forecasts

Patterns learned from other products help with short histories, too — without training a new model for each product.

02Understand

Your promotions and prices feed into the forecast.

The model has learned from millions of products how demand develops – with trends, seasonality and product life cycles. It combines that knowledge with your sales history and the most important driving factors. That is your forecast.

03Forecast

Uncertainty grows with demand.

Demand is rising, and past promotions have produced different responses. Q‑Foundry considers the history and the next planned promotion date. In this example, expected demand rises on that day. The size of the increase remains uncertain, so the forecast band widens.

Demand and promotion planningUnits / day
90% forecast band · Q05 to Q95 Expected demand Selected quantile · Q90

For well-calibrated forecasts, the 90% band describes the range containing actual demand on about 9 out of 10 days over time.

Q90 is selected for planning here: according to the model, demand stays below it in 90% of cases. The band extends above this curve — higher demand remains possible.

04Plan

Every constraint, optimized together.

A larger order can reduce ordering costs while tying up more capital and storage space. Our mathematical optimization considers these trade-offs and all specified constraints simultaneously — across products and dates.

Many requirements. One joint optimization.

Demand & availability

  • Demand quantiles
  • Target availability
  • Inventory & open orders
  • …

Supplier rules

  • Minimum quantities & pack sizes
  • Lead times & delivery days
  • Price tiers
  • …

Operating limits

  • Storage capacity
  • Budget & liquidity
  • Shelf life & remaining life
  • …

Trade-offs

  • Availability ↔ tied-up capital
  • Volume discounts ↔ expiry risk
  • Ordering costs ↔ holding costs
  • …
Optimization

The model considers all requirements simultaneously and determines order quantities at the lowest total cost.

Total costs

i: product · τ: date · q: quantity · I: inventory

  • Target availability

    Provides supply at the selected service level.

  • Price tiers

    Create steps in the cost surface.

  • Minimum quantities

    Exclude orders below the minimum.

  • Pack sizes

    Allow only discrete quantities.

Total cost C Feasible combinations

Order proposal q*

At the lowest total cost

  • Product A · xₐ units
  • Product B · xᵦ units
  • …

Two quantity axes show a slice of the multidimensional problem. The surface represents costs; points mark feasible quantity combinations. Further products and dates add dimensions.

Products share suppliers, storage space and budget. An ordering decision therefore changes the options available for other products, too.

The foundation

From research.
Into practice.

Our technology grew out of our research in the Data Driven Decisions (D3) group at Julius-Maximilians-Universität Würzburg. As a university spin-off, SOLISTIQ brings this research into practice.

Published

  1. ECIS 2025

    What Can We Learn from LLMs? Building a Foundation Model for Inventory Management

    Maichle, Stein, Pibernik

  2. NeurIPS 2024 · Workshop

    In-context Quantile Regression for Multi-product Inventory Management using Time-series Transformers

    Maichle, Mukherjee, Günder, Antonov, Stein, Pibernik

  3. ICLR 2025 · Workshop

    Predictive Inference Is Really Free with In-Context Learning

    Mukherjee, Antonov, Günder, Maichle