2025技术雷达_AI推理_机器学习编排与代理型AI工具及平台研究报告_41页_1mb

报告摘要

CNCF Technology Radar Summary: AI Inferencing, ML Orchestration, and Agentic AI Tools

Core Content

This report presents the results of a survey conducted in Q3 2025, where 300+ professional developers associated with cloud native development were asked about their experience and opinions on AI inferencing tools and engines, ML orchestration tools, and agentic AI platforms, projects, and systems. The technologies were selected by the Cloud Native Computing Foundation (CNCF) and its End User Community based on relevance and importance. The findings provide insights into which AI/ML tools are gaining traction and how they are being perceived in terms of maturity, usefulness, and likelihood to recommend.

Key Insights

Adopt Position Technologies:
- AI Inferencing: NVIDIA Triton, DeepSpeed, TensorFlow Serving, and BentoML.
- ML Orchestration: Airflow and Metaflow.
- Agentic AI: Model Context Protocol (MCP) and Llama Stack.
Top Performers:
- NVIDIA Triton received the highest ratings for both maturity and usefulness, with 50% and 41% of developers giving it 5-star ratings, respectively.
- Adlik received the most recommendations, with 92% of users recommending it.
- Metaflow received the highest maturity ratings, while Airflow was the most likely to be recommended and had the highest usefulness rating.
- MCP leads in maturity and usefulness, but Agent2Agent has the highest recommendation score at 94%.
Technology Categorization:
- Technologies are categorized into Adopt, Trial, Assess, and Hold based on their usage, usefulness, maturity, and recommendation scores.
- Adopt technologies are considered reliable and widely applicable.
- Trial technologies are worth exploring for specific use cases.
- Assess technologies require further evaluation.
- Hold technologies are less mature or useful.
CNCF Maturity Model:
- The report’s radar positions do not directly correlate with the CNCF maturity model (Sandbox, Incubating, Graduated), which reflects different stages of development and adoption.

AI Inferencing Tools and Engines

Maturity

NVIDIA Triton has the highest maturity ratings, with 50% of users giving it a 5-star rating and 30% a 4-star rating.
LMCache is the second-highest with 43% of 5-star ratings but fewer 4-star ratings (21%).
Ollama is divisive, with 34% 5-star ratings and 23% 1- and 2-star ratings, indicating a mixed perception.
Envoy AI Gateway and llama.cpp received the highest proportion of negative ratings (18% and 15%, respectively), with Envoy AI Gateway having a much lower proportion of positive ratings.

Usefulness

NVIDIA Triton leads in usefulness, with 41% of developers giving it 5-star ratings and 38% 4-star ratings.
DeepSpeed and BentoML also have strong usefulness ratings, with 35% and 38% of 5-star ratings, respectively.
Flyte and Seldon Core show lower usefulness ratings, with Flyte receiving 68% of 4- and 5-star ratings and Seldon Core 55%.

Recommendation

NVIDIA Triton has the highest likelihood to recommend (57%).
Adlik and Seldon MLServer have the highest recommendation scores (92% each), despite lower maturity and usefulness ratings.
BentoML has a high recommendation score (84%) but lower likelihood to recommend (33%).

Machine Learning (ML) Orchestration Tools

Maturity

Feast leads in maturity with 46% of 5-star ratings, but Metaflow and Argo Workflows have higher combined 4- and 5-star ratings (84% and 80%, respectively).
Flyte has a lower proportion of 4- and 5-star ratings (47%), with 44% giving it a 3-star rating, suggesting it fails to impress rather than being poorly received.

Usefulness

Metaflow, Airflow, and Feast all have high 5-star ratings (43% each).
Airflow has no negative ratings, indicating consistent satisfaction.
Flyte and Seldon Core show lower usefulness ratings, with Flyte having 68% of 4- and 5-star ratings and Seldon Core 55%.

Recommendation

Metaflow has the highest likelihood to recommend (51% highly likely, 35% likely).
Airflow and Argo Workflows have 90% cumulative likelihood to recommend.
BentoML has a high recommendation score (84%) but lower likelihood to recommend (33%).

Agentic AI Platforms, Projects, and Systems

Maturity

Agent2Agent and Llama Stack have the highest maturity ratings, with 38% and 35% of 5-star ratings, respectively.
MCP, despite being in the adopt position, has a slightly lower 5-star rating (33%) but the highest combined 4- and 5-star ratings (73%).

Usefulness

Autogen has the highest proportion of 5-star ratings (45%), while MCP has the highest combined 4- and 5-star ratings (80%).
MCP is seen as broadly useful, while Autogen is popular in a more specialized community.

Recommendation

Agent2Agent has the highest recommendation score at 94%, with 48% likely and 46% highly likely to recommend it.
MCP is also highly recommended, with 84% of users indicating they would recommend it.

Conclusion

The AI/ML tooling landscape is showing a clear maturity gradient, with established technologies like NVIDIA Triton, Airflow, and MCP in the adopt position due to their reliability and broad utility.
Emerging solutions are demonstrating continued innovation in areas like agent-based architectures and standardized integration protocols.
CNCF projects are playing a crucial role in the development lifecycle of AI/ML technologies, from experimental to production-ready.
The survey highlights that cloud native approaches are essential for AI/ML workloads, even if developers do not explicitly identify their workflows as cloud native.
41% of ML.AI developers are currently categorized as cloud native, and this number is expected to increase.

Methodology

Likert Scales: Used to capture developers' opinions on maturity and usefulness (1 to 5 stars).
Subjectivity: While subjective, these ratings offer valuable insights into developer perceptions and experiences.
Demographics: Respondents were recruited from third-party panels, with no organization-specific data included for privacy.
Industry Involvement: The survey included developers from various sectors, with a focus on cloud native development approaches.

展开完整摘要