CNCF技术雷达：AI推理、机器学习编排与代理型AI工具及平台

AI INFERENCING,ML ORCHESTRATION, Can I share data from this report? 1. License Grant 2. Limitation of Liability SlashDataTM, believes the statements contained in this publication to be based upon This report is licensed undertheCreative Commons Attribution-NoDerivativesLicence 4.0 (International). Put simply, subject to the terms and conditions of thislicense, you are free to: information that we consider reliable, but we do not represent that it is accurate orcomplete and it should not be relied upon as such. Opinions expressed are currentopinions as of the date appearing in this publication only and the information,including the opinions contained herein, are subject to change without notice. Use ofthis publication by any third party for whatever purpose should not and does not Share—You can reproduce the report or incorporate parts of the report into one ormore documents or publications, for commercial and non-commercial purposes. Under the following conditions: Attribution—You must give appropriate credit toSlashDataTM, and to the CloudNative Computing Foundation as sponsors of this report, and indicate if changeswere made. In that case, you may do so in any reasonable manner, but not in any way SlashDataTM, its affiliates, and representatives shall have no liability for any direct,incidental, special, or consequential damages or lost profits, if any, suffered by any NoDerivatives—you cannot remix or transform the content of the report. You maynot distribute modified content. Liam Bollmann-Dodd Senior Market Research Consultant Liam is a former experimental antimatterphysicist, and he obtained a PhD in Physicswhile working at CERN. He is interested inthe changing landscape of cloud Álvaro Ruiz Cubero liam.dodd@slashdata.co Research Manager Álvaro is a market research analyst with abackground in strategy and operationsconsulting. He holds a Master’s in BusinessManagement and believes in the power of data- ABOUT THE AUTHORS alvaro.ruiz@slashdata.co KEY INSIGHTS •For AI inference tools, NVIDIA Triton,DeepSpeed,TensorFlow Serving, andBentoMLwere the projects thatdevelopers cumulatively placed in the adopt position.➜ •Metaflowreceived the highest maturity ratings, whileAirflow was the most likely to be recommended andreceived the highest usefulness rating.➜ •NVIDIA Triton received the highest ratings for maturity and •For machine learning orchestration,BentoMLwas placed inthe trial position, indicating that technologies crossingmultiple use cases can still succeed in each area but arelikely to struggle to be market leaders in all.➜ •Adlikreceived the most recommendations, with 92% ofcurrent or former users recommending it to otherdevelopers.➜ •Model Context Protocol (MCP) and Llama Stack are theagentic AI projects that developer perception placed in the •Airflow andMetaflowwere the two technologies that rose •MCP leads on maturity and usefulness ratings, butAgent2Agent was the most likely to be recommended by its 1. Introduction In Q3 2025, more than 300 professional developers usingtechnologies associated with cloud native development wereasked about their experience and opinions with regard to AIinference tools and engines, machine learning (ML) orchestrationtools, and agentic AI platforms, projects, and systems. The For the products or tools they were familiar with, they ratedthem on their usefulness and maturity and indicated how likelythey were to recommend that technology to other developers.Within the context of this report, usefulness was defined as how 1. Introduction Note:These radar positions do not necessarily correlate with the CNCFmaturity model (Sandbox, Incubating, and Graduated), whichcorresponds to the Innovators, Early Adopters, and Early Majority tiers Based on the usage, usefulness and maturity ratings, and how likely they are to recommend a given technology, wecategorized the technologies into four groups: adopt, trial,assess, and hold. 'Adopt' technologies are considered reliablechoices for most use cases, while 'trial' technologies are worthexploring to see if they meet your specific needs. 'Assess'technologies require careful evaluation before committing, and Sandbox:Sandbox projects are in their earliest stages, meant forexperimentation and foundational growth. They are newer technologiesthat represent initial concepts and technologies with significant room Incubating:Projects that have a solidified technical vision and agrowing contributor base but are still maturing in terms of community Graduated:Graduated projects are widely adopted and reliable. Theyhave established a diverse community base supported by maturetechnical policies and governance. AI INFERENCINGTOOLS AND 2. AI Inferencing Tools and Engines For AI inferencing tools and engines, we find NVIDIATriton,DeepSpeed, TensorFlow Serving, andBentoMLasthe technologies that respondents would cumulativelyplace in the ‘adopt’ position of the technology radar. Developers familiar with AI