World Summit AI Amsterdam | 7-8 October 2026

Qluon at World Summit AI

Startup Accelerator · Booth S16

Qluon builds supervisory control for AI training. Optimizer+ works with existing AdamW, Muon, and Lion optimizers, using live training signals to adjust optimizer behavior and runtime execution. The aim is lower GPU-hours while protecting model quality.

Arrange a meeting at S16 Build a pilot brief

54 runsMatched Matrix54 breadth study

15.2-41.9%Lower GPU-hours than tested accelerated runtimes

~1/3 GPU-hMuon in a separate one-seed three-pass study

At booth S16

Bring one workload. Leave with a testable next step.

Show us your training setup, quality target, and compute cost. We will map them to the evidence, discuss how Optimizer+ fits around your existing stack, and decide whether a private pilot is worthwhile. Your team keeps its model, data, and environment; the pilot compares model quality, GPU-hours, and operational overhead against your accepted baseline.

Bring these five facts

A useful conversation starts from your baseline.

Workload: PT, CPT, or SFT
Model: architecture, scale, and adaptation method
Optimizer: current family and accepted tuning
Objective: quality metric and fixed budget
Cost: hardware, runtime, and GPU-hours

Leave with four decisions

No open-ended proof of concept.

Comparator: your accepted native baseline
Gate: quality, GPU-hours, and reliability targets
Boundary: integration, fallback, and data controls
Stop rule: pass, conditional pass, or stop

Evidence boundary: Matrix54 is a two-seed internal Qwen2.5-3B LoRA CPT study; its reported GPU-hour range compares with tested accelerated runtimes. The separate 18-run, three-pass study uses one seed on Qwen2.5-1.5B LoRA CPT. Neither establishes a universal gain; customer replication uses your accepted baseline.

A measurable next step

Use your workload, your baseline, and your success gate.

Qluon supplies the control layer and evidence protocol. Your team keeps its model, data, and environment.

Define the pilot