Overview

XReduce is the measurement layer for intelligent systems. We capture how self-hosted AI workloads behave under real execution conditions - power, latency, memory, throughput, quality outcomes - and turn that signal into durable optimization.

Behavior under real execution conditions is what makes intelligent systems comparable, understandable, and ultimately improvable. XReduce measures what's actually happening on your infrastructure, against your workloads, and produces the evidence you need to keep your models and your systems running well together.

What you get

• Multi-dimensional capture. Power, latency, memory, throughput, and quality outcomes - per workload, per model, per request.
• A coherent picture of your system. Workload signatures, cost-quality matrices, hardware-aware analysis, comparable across runs.
• Decisions you can defend. Which model for which workload, paired with full-system context, updated as evidence accumulates.

The workflow

XReduce is a CLI-driven workflow. From a clean machine to your first evaluation results is five commands:

# 1. Set up the machine - installs Python 3.11, Node, the CLI Skill, deps
curl -fsSL https://www.xreduce.ai/install.sh | bash

# 2. Authenticate
xreduce login

# 3. Scaffold a project folder
xreduce init my-org/my-model

# 4. Register the model with XReduce
xreduce model create --name my-org/my-model --task-type text-to-sql

# 5. Sanity check + run the evaluation
xreduce profile --config my-org/my-model/config.yaml
xreduce evaluate --config my-org/my-model/config.yaml

Each step has its own walkthrough in the sidebar. Account and Authentication covers login and credentials. Set up a model covers init, model create, and benchmark data. Run XReduce covers profile, evaluate, and compare.

Requirements

• Python 3.11 (the SDK ships cp311 wheels only)
• Linux or macOS
• A GPU if you intend to profile/evaluate locally