Scroll Down

The Infrastructure forEvolutionary AI Systems

Enabling intelligence to adapt, create, and evolve across digital and physical worlds.

Get StartedGitHub

EvoX is All You Need

01Why Evolution

The Power of Evolution

Evolution is nature's most proven way to create intelligence. EvoX brings that loop to AI: generate many candidates, test them for real, keep what works, and build on it — so capability compounds through experiment instead of hand-tuning.

  1. 01Generate
  2. 02Execute
  3. 03Evaluate
  4. 04Select
  5. 05Improve

02Applications

Evolution Across AI Paradigms

From models to agents, robots, and scientific discovery — one evolutionary loop, applied wherever intelligence needs to adapt.

01

Generative AI

Evolving the models themselves — merging expert models, optimizing generation, and searching new architectures.

EvoGM · Evolutionary model merging
02Featured

Agentic AI

Recursive agents that decompose, build, and evolve entire codebases — from a blank repository to shipped software.

Genesis · Recursive agentic system
03

Physical AI

Evolving robot control policies through massively parallel interaction with simulated environments.

EvoRL · Evolutionary reinforcement learning
04

AI for Science

Data-driven black-box optimization for scientific and engineering design, validated on supercritical airfoils.

EvoGO · Supercritical airfoil design

03Hardware

Computing Without Boundaries

From a laptop to a GPU server, the same EvoX code runs tensorized across heterogeneous hardware.

  • NVIDIA GPU

    Supported

  • AMD GPU

    Supported via ROCm

  • Vulkan

    Supported via EvoX Compiler

  • Apple Silicon

    Supported via MPS

  • Huawei Ascend

    Project-level validation

  • CPU

    Supported

  • Distributed

    Experimental

80.2×

Median GPU-over-CPU speedup across 276 algorithm comparisons as problem dimensionality scales

Read the study

Up to 1,113×

Peak GPU speedup of tensorized multiobjective algorithms over their CPU counterparts

Read the study

Status reflects the validated scope. Huawei Ascend results come from specific project validations and do not imply full-framework compatibility.

04EvoX

Explore EvoX

01

GenesisFeatured Project

Watch software evolve autonomously

02

Ecosystem

Libraries for optimization, RL, GP, NAS, and more

03

Documentation

Guides, tutorials, and API reference

04

GitHub

Star, fork, and contribute to the source

05Blog

Latest News

View All News
Releases

EvoX v1.4.0 Release Note

EvoX v1.4.0 is a maintenance release focused on bug fixes, compatibility improvements and polish across the framework, alongside the announcement of EvoX Genesis.

News

EvoX Genesis Completes WASM Render in About Two Days for $36

Using DeepSeek V4 Flash, EvoX Genesis completed WASM Render in about two days for just $36 in model charges, becoming the first AI system worldwide to submit a result for the challenge. Team test records show 2,017/2,019 Khronos CTS tests passed (99.90%), including all 1,184/1,184 WebGL2.0 conformance tests run.

News

EvoCoCo: Automatic MATLAB-to-EvoX Code Conversion for GPU-Accelerated Multiobjective Optimization

EvoCoCo converts MATLAB multiobjective evolutionary algorithms into GPU-accelerated EvoX implementations through semantics-guided tensorization, execution feedback, and multi-agent validation. Evaluated on 48 algorithms, it achieves 88.2% optimization-fidelity coverage.