Intelligence BooM

Overview

Intelligence BooM is an open source AI foundational software solution, the result of a joint effort by community members and partners. It comprises over 20 open source components and the 7 major platforms: heterogeneous convergence, task management, data management, runtime acceleration, domain model platform, intelligent application platform, and full-stack security. Users can apply the reference implementation to release their commercial applications, contribute to community development, and engage in technological and agent innovation.

Vision

Break Technical Barriers

Democratize LLM inference through full-stack open source innovation.

Fuel Industry Transformation

Build a collaborative AI application platform to drive industrial intelligence.

Advance Ecosystem Collaboration

Forge universal standards and collaboration across the community, academia, and enterprises.

Intelligence BooM 25.11 (Qiaoyu Noodle)

Empower domain-specific AI with the pioneering OS-level fine-tuning model and seamless NPU orchestration

Try Now

Peak Performance

High-throughput, low-cost inference via convergence and allocation of heterogeneous computing power

50%+ lower TTFT for multi-turn and long-context dialogues via multi-tier caching

Effortless Development

AIKG for Ascend: production-ready performance and 5x+ faster Triton operator development

Full-stack Capability

Cost-efficient fine-tuning with Atlas 300 series and 50+ pre-trained models

Seamless Dify integration for out-of-the-box agentic workflows

Hassle-free Experience

< 15 minutes from image to first loss

Decoupled MindSpore graph mode for PyTorch

Ecosystem Building

Co-launched vLLM-omni and LMCache-Ascend through upstream partnerships

openEuler OS Model: 671B-equivalent performance in system tuning with a 4B footprint and CPU inference

Intelligent application platform

openEuler Intelligence

Intelligent tuning

Intelligent O&M

Intelligent question answering

Deep research

Dify

Langchain

...

Domain model platform

openEuler OS Model

...

Runtime acceleration platform

vLLM

SGLang

LLaMA Factory

...

...

PyTorch

...

MS-InferRT

Data management platform

openGauss

Task management platform

openYuanrong

RAY

openFuyao

K8S

oeDeploy

...

Heterogeneous convergence platform

Heterogeneous Convergence Compilation

AscendNPUIR

AKG

Heterogeneous Fusion Operating System Kernel

FalconFS

GMEM

Xsched

(user-space)

Xsched

(kernel-space)

ModelFS

xMig

CPU

NPU

GPU

Intelligence BooM 25.07 (Hui Noodle)

Build out-of-the-box inference + agentic workflow solutions

Try Now

Versatile Compatibility

Next-gen tuning and O&M agents powered by RAG

Unified support for diverse hardware, including x86/Arm CPUs, Ascend NPUs, and NVIDIA GPUs

Dependable Security

Trusted platform foundation with robust protection for data and model privacy

Seamless Onboarding

Effortless distributed deployment in 15 minutes via oeDeploy with a single command

50+ mainstream open source models supported with 0-day adaptation

Flexible scaling for DeepSeek-V3/R1 from single-node prototyping to massive-scale cluster production

CPU+xPU collaborative inference with flexible MoE scaling and expert offloading

Streamlined Integration

Enterprise AI agent development platform with enhanced workflow integration capabilities

Streamlined operator development based on Triton and SWFT

Multi-agent LLMs for automated operator generation across diverse frontends and backends