语言 & 框架
采纳
-
85. Fastify
We continue to have a positive experience with Fastify — a fast, unopinionated, low-overhead web framework for Node.js. It provides all the essential capabilities of a minimal web framework — including parsing, validation and serialization — along with a robust plugin system and strong community support. Our teams have seen no significant downsides in using Fastify over alternatives such as Express.js, while also gaining measurable performance improvements, making it a compelling choice for minimal web development in Node.js.
-
86. LangGraph
LangGraph is an orchestration framework for building stateful multi-agent applications using LLMs. It provides low-level primitives such as nodes and edges, along with built-in features that give developers granular control over agent workflows, memory management and state persistence. This means developers can start with a simple pre-built graph and scale to complex, evolving agent architectures. With support for streaming, advanced context management and resilience patterns like model fallbacks and tool error handling, LangGraph enables you to build robust, production-grade agentic applications. Its graph-based approach ensures predictable, customizable workflows and simplifies debugging and scaling. Our teams have had strong results using LangGraph to build multi-agent systems thanks to its lightweight and modular design.
-
87. vLLM
vLLM is a high-throughput, memory-efficient inference engine for LLMs that can run in the cloud or on-premises. It supports multiple model architectures and popular open-source models. Our teams deploy dockerized vLLM workers on GPU platforms such as NVIDIA DGX and Intel HPC, hosting models including Llama 3.1 (8B and 70B), Mistral 7B and Llama-SQL for developer coding assistance, knowledge search and natural language database interactions. vLLM is compatible with the OpenAI SDK standard, enabling consistent model serving. Azure's AI Model Catalog uses a custom inference container built on vLLM to boost serving performance, with vLLM as the default inference engine due to its high throughput and efficient memory management. The vLLM framework has become a go-to framework for large-scale model deployments.
试验
-
88. Crossplane
Since its last appearance on the Radar, Crossplane adoption has continued to grow, particularly for extending Kubernetes clusters. In our work, we've found Crossplane excels in specific use cases rather than as a general-purpose infrastructure-as-code (IaC) tool. Our earlier observations still hold true: Crossplane works best as a companion to workloads deployed within Kubernetes, not as a full replacement for tools like Terraform. Teams that went "all-in" on Crossplane as their primary IaC solution often struggled, whereas those using it pragmatically — for targeted, custom use-cases — saw strong results. Offloading resource lifecycle management to Crossplane while using XRD APIs for lightweight customization has proven especially effective. Crossplane is particularly valuable when managing resources whose lifecycles are straightforward but not native to Kubernetes. While it can now create Kubernetes clusters — a capability that was previously missing — we advise caution when adopting Crossplane as a complete Terraform replacement. In our experience, it works best when IaC serves as the foundational layer, with Crossplane layered on top for specialized requirements.
-
89. DeepEval
DeepEval is an open-source, Python-based evaluation framework for assessing LLM performance. It can be used to evaluate retrieval-augmented generation (RAG) and other applications built with frameworks such as LlamaIndex or LangChain, as well as to baseline and benchmark models. DeepEval goes beyond word-matching scores, assessing accuracy, relevance and consistency to provide more reliable evaluation in real-world scenarios. It includes metrics such as hallucination detection, answer relevancy and hyperparameter optimization and supports GEval for creating custom, use case–specific metrics. Our teams are using DeepEval to fine-tune agentic outputs using the LLM as a judge technique. It integrates with pytest and CI/CD pipelines, making it easy to adopt and valuable for continuous evaluation. For teams building LLM-based applications in regulated environments, Inspect AI, developed by the UK AI Safety Institute, offers an alternative with stronger focus on auditing and compliance.
-
90. FastMCP
The Model Context Protocol (MCP) is rapidly becoming a standard for providing context and tooling for LLM applications. However, implementing an MCP server typically involves substantial boilerplate for setup, protocol handling and error management. FastMCP is a Python framework that simplifies this process by abstracting away protocol complexity and allowing developers to define MCP resources and tooling through intuitive Python decorators. This abstraction enables teams to focus on business logic, resulting in cleaner, more maintainable MCP implementations.
While FastMCP 1.0 is already incorporated into the official SDK, the MCP standard continues to evolve quickly. We recommend monitoring the 2.0 release and ensuring teams stay aligned with changes to the official specification.
-
91. LiteLLM
LiteLLM is an SDK that provides seamless integration with multiple LLM providers through a standardized OpenAI API format. It supports a wide range of providers and models, offering a unified interface for text completion, embeddings and image generation. By abstracting provider-specific API differences, LiteLLM simplifies integration and automatically routes requests to the correct model endpoint. It also includes production-grade features such as guardrails, caching, logging, rate limiting and load balancing through its proxy framework. As organizations embed AI-powered applications more deeply into workflows, governance and observability become essential. Our teams have been using LiteLLM as an AI gateway to standardize, secure and gain visibility into enterprise-wide AI usage.
-
92. MLForecast
MLForecast is a Python framework and library for time series forecasting that applies machine learning models to large-scale data sets. It simplifies the typically complex process of automated feature engineering — including lags, rolling statistics and date-based features — and is one of the few libraries with native support for distributed computing frameworks like Spark and Dask, ensuring scalability. It also supports probabilistic forecasting using methods such as conformal prediction, providing quantitative measures of forecast uncertainty. In our evaluation, MLForecast scaled efficiently to millions of data points and consistently outperformed comparable tools. For teams looking to rapidly operationalize time series forecasting on high-volume data, MLForecast is a compelling choice.
-
93. Nuxt
Nuxt is an opinionated meta-framework built on top of Vue.js for creating full-stack web applications, often known as the "Next.js for Vue.js." Similar to its React counterpart, Nuxt provides SEO-friendly capabilities such as pre-rendering, server-side rendering (SSR) and metadata management, making it a strong choice for building performance-oriented, SEO-optimized websites on the Vue.js stack.
Nuxt is backed by Vercel, the same company behind Next.js, and supported by a strong community and ecosystem of official and third-party modules. These modules simplify integration of features such as image processing, sitemaps and Tailwind CSS. Nuxt is a good choice for teams looking for a comprehensive, opinionated framework for building SEO-friendly applications with Vue.js.
-
94. Phoenix
We continue to have positive experiences with Phoenix, a server-side web MVC framework written in Elixir. Phoenix builds on the rapid application development and developer experience learnings of Ruby on Rails, while also advancing into functional programming paradigms.
In this volume, we’re highlighting the release of Phoenix LiveView 1.0. LiveView is an HTML-over-the-wire solution — similar to HTMX or Hotwire — that enables developers to build rich, real-time user experiences entirely with server-rendered HTML. While similar technologies usually handle partial updates through HTML switching, LiveView provides a full component-based architecture via LiveComponents, offering composition, props passing, state management and lifecycle hooks akin to React or Vue.js. LiveView combines the productivity and scalability of Phoenix with robust complexity management, making it well-suited for building highly interactive frontends without the need for a full JavaScript framework.
-
95. Presidio
Presidio is a data protection SDK for identifying and anonymizing sensitive data in structured and unstructured text. It detects personally identifiable information (PII) such as credit card numbers, names and locations using named entity recognition, regular expressions and rule-based logic. Presidio supports custom entity recognizers and de-identification pipelines, allowing organizations to tailor it to their privacy and compliance needs. Our teams have used Presidio in enterprise environments with strict data-sharing controls when integrating with LLMs. While it automates the detection of sensitive information, it isn’t foolproof and may miss or misidentify entities. Teams should exercise caution when relying on its results.
-
96. Pydantic AI
Pydantic AI continues to prove itself as a stable, well-supported, open-source framework for building GenAI agents in production. Built on the trusted Pydantic foundation, it offers strong type safety, first-class observability through OpenTelemetry and built-in evaluation tooling. The release of version 1.0 on September 4, 2025 marked a significant milestone in its maturity. Since then, we’ve found it reliable and widely adopted for its simplicity and maintainability, joining the ranks of other popular agent frameworks like LangChain and LangGraph.
Recent updates have made it easier to implement Model Context Protocol (MCP) servers and clients, with added support for emerging standards such as AG-UI and A2A. With its clean API and growing ecosystem, Pydantic AI has become a compelling choice for our teams building production-ready GenAI applications in Python.
-
97. Tauri
Tauri is a framework for building high-performance desktop applications using a single web UI codebase. Unlike traditional web wrappers such as Electron, Tauri is built on Rust and leverages the operating system's native webview, resulting in smaller binaries and stronger security. We first evaluated Tauri several years ago; since then, it has expanded beyond desktop to support iOS and Android. The latest version introduces a more flexible permission and scope model, replaces the older permission list and features a hardened inter-process communication (IPC) layer that supports raw data transfer and improves performance. These updates are backed by an external security audit. Together with official distribution guidelines for major app stores, these improvements further strengthen Tauri's position in the cross-platform development space.
评估
-
98. Agent Development Kit (ADK)
Agent Development Kit (ADK) is a framework for developing and deploying AI agents that applies modern software engineering discipline rather than relying solely on prompting. It introduces familiar abstractions such as classes, methods, workflow patterns and CLI support. Compared to frameworks like LangGraph or CrewAI, ADK’s strength lies in its deep integration with Google’s AI infrastructure, providing enterprise-ready grounding, data access and monitoring. It’s also designed for interoperability, supporting tool wrappers and the A2A protocol for agent-to-agent communication. For organizations already invested in GCP, ADK presents a promising foundation for building scalable, secure and manageable agentic architecture. Though still early in its evolution, it signals Google’s direction toward a native, full-stack agent development environment. We recommend keeping a close eye on its maturity and ecosystem growth.
-
99. Agno
Agno is a framework for building, running and managing multi-agent systems. It offers the flexibility to create fully autonomous agents or controlled, step-based workflows, with built-in support for human-in-the-loop, session management, memory and knowledge. We appreciate its focus on efficiency, with impressive agent startup times and low memory consumption. Agno also comes with its runtime, AgentOS, a FastAPI application with an integrated control plane for streamlined testing, monitoring and management of agentic systems.
-
100. assistant-ui
assistant-ui is an open-source TypeScript and React library for AI chat interfaces. It handles the complex parts of chat UI implementation — such as streaming, state management for message editing and branch switching and common UX features — while allowing developers to design their own components using Radix primitives. It supports integration with popular runtimes, including Vercel AI SDK and LangGraph, and offers customizable runtime solutions for complex use cases. We’ve successfully built a simple chat interface with assistant-ui and have been pleased with the results.
-
101. AutoRound
Intel's AutoRound is an advanced quantization algorithm for compressing large AI models, such as LLMs and vision language models (VLMs), with minimal loss of accuracy. It reduces model size to ultra-low bit widths (2–4 bits) using sign-gradient descent optimization and applies mixed bit widths across layers for optimal efficiency. This quantization process is also remarkably fast: You can quantize a 7-billion-parameter model in just minutes on a single GPU. Since AutoRound integrates with popular inference engines such as vLLM and Transformers, it's an attractive option for quantizing models.
-
102. Browser Use
Browser Use is an open-source Python library that enables LLM-based agents to operate web browsers and interact with web applications. It can navigate, enter data, extract text and manage multiple tabs to coordinate actions across applications. The library is particularly useful when AI agents need to access, manipulate or retrieve information from web content. It supports a range of LLMs and leverages Playwright to combine visual understanding with HTML structure extraction for richer web interactions. Our teams integrated Browser Use with the Pytest framework and Allure reporting to explore automated testing with LLMs. Test steps were written in natural language for the agent to execute, capturing screenshots on assertions or failures. The goal was to enable off-hours QA by automatically pulling test cases from Confluence for post-development verification. Early results are promising, though the agent's post-task responses often lack detailed failure descriptions, so require custom error reporting.
-
103. DeepSpeed
DeepSpeed is a Python library that optimizes distributed deep learning for both training and inference. For training, it integrates technologies such as the Zero Redundancy Optimizer (ZeRO) and 3D parallelism to efficiently scale models across thousands of GPUs. For inference, it combines tensor, pipeline, expert and ZeRO parallelism with custom kernels and communication optimizations to minimize latency. DeepSpeed has powered some of the world's largest language models, including Megatron-Turing NLG (530B) and BLOOM (176B). It supports both dense and sparse models, delivers high system throughput and allows training or inference across multiple resource-constrained GPUs. The library integrates seamlessly with popular Hugging Face Transformers, PyTorch Lightning and Accelerate, making it a highly effective option for large-scale or resource-limited deep learning workloads.
-
104. Drizzle
Drizzle is a lightweight TypeScript ORM. Unlike Prisma ORM, it gives developers both a simple SQL-like API and a more traditional ORM-style query interface. It also supports extracting schemas from existing databases, enabling both database-first and code-first approaches. Drizzle was designed with serverless environments in mind: It has a small bundle size and supports prepared statements, allowing SQL queries to be precompiled so that the database driver executes binary SQL directly instead of parsing queries each time. Its simplicity and serverless support make Drizzle an appealing choice for ORM use in the TypeScript ecosystem.
-
105. Java post-quantum cryptography
Quantum computers continues to advance rapidly, with SaaS offerings like AWS Braket now providing access to quantum algorithms across multiple architectures.
Since March, Java 24 has introduced Java post-quantum cryptography, adding support for post-quantum cryptographic algorithms such as ML-KEM and ML-DSA, and .Net 10 has expanded its support as well. Our advice is simple: if you’re building software in these languages, begin adopting quantum-safe algorithms now to future-proof your systems.
-
106. kagent
Kagent is an open-source framework for running agentic AI inside Kubernetes clusters. It enables LLM-based agents to plan and execute operational tasks such as diagnosing issues, remediating configurations or interacting with observability tools through Kubernetes-native APIs and Model Context Protocol (MCP) integrations. Its goal is to bring "AgentOps" to cloud-native infrastructure by combining declarative management with autonomous reasoning.
As a CNCF Sandbox project, Kagent should be introduced with care, particularly given the risks of granting LLMs operational management capabilities. Techniques such as toxic flow analysis can be especially valuable when assessing and mitigating these risks.
-
107. LangExtract
LangExtract is a Python library that uses LLMs to extract structured information from unstructured text based on user-defined instructions. It processes domain-specific materials — such as clinical notes and reports — identifying and organizing key details while keeping each extracted data point traceable to its source. The extracted entities can be exported as a
.jsonlfile, a standard format for language model data and visualized through an interactive HTML interface for contextual review. Our teams evaluated LangExtract for extracting entities to populate a domain knowledge graph and found it effective for transforming complex documents into structured, machine-readable representations. -
108. Langflow
Langflow is an open-source, low-code platform for building and visualizing LLM workflows. Built on top of LangChain, it allows developers to chain prompts, tools, vector databases and memory components through a drag-and-drop interface, while still supporting custom Python code for advanced logic. It's particularly useful for prototyping agentic applications without writing full back-end code. However, Langflow is still relatively new and has some rough edges for production use. Our usual caution around low-code platforms applies here. That said, we like that Langflow's workflows can be defined and versioned as code, which can mitigate some of the drawbacks of low-code platforms.
-
109. LMCache
LMCache is a key-value (KV) cache solution that accelerates LLM serving infrastructure. It acts as a specialized caching layer across a pool of LLM inference engines, storing precomputed KV cache entries for texts likely to be processed multiple times, such as chat histories or document collections. By persisting these values on disk, prefill computations can be offloaded from the GPU, reducing time-to-first-token (TTFT) and cutting inference costs across demanding workloads like RAG pipelines, multi-turn chat applications and agentic systems. You can integrate LMCache with major inference servers such as vLLM or NVIDIA Dynamo and think it’s worth assessing its impact on your setup.
-
110. Mem0
Mem0 is a memory layer designed for AI agents. Naive approaches often store entire chat histories in a database and reuse them in future conversations, which leads to excessive token usage. Mem0 replaces this with a more sophisticated architecture that separates memory into short-term recall and an intelligent long-term layer that extracts and stores only salient facts and relationships. Its architecture combines a vector store for semantic similarity with a knowledge graph for understanding temporal and relational data. This design significantly reduces context token usage while enabling agents to maintain long-term awareness, which is extremely useful for personalization and many other use cases.
-
111. Open Security Control Assessment Language (OSCAL)
The Open Security Controls Assessment Language (OSCAL) is an open, machine-readable information exchange format designed to increase automation in compliance and risk management, and help teams move away from text-based manual approaches. Led by the National Institute of Standards and Technology (NIST), OSCAL provides standard representations in XML, JSON and YAML for expressing security controls associated with industry frameworks such as SOC 2 and PCI, as well as government frameworks such as FedRAMP in the United States, Singapore's Cybersecurity Control Catalogue and Australia's Information Security Manual.
While OSCAL has not yet been widely adopted outside the public sector and its ecosystem is still maturing, we’re excited by its potential to streamline security assessments, reduce reliance on spreadsheets and box-ticking exercises and even enable automated compliance when incorporated into compliance-as-code and continuous compliance platforms.
-
112. OpenInference
OpenInference is a set of conventions and plugins; it’s complementary to OpenTelemetry and designed to observe AI applications. It provides standardized instrumentation for machine-learning frameworks and libraries, which helps developers trace LLM invocations along with surrounding context such as vector store retrievals or external tool calls to APIs and search engines. Spans can be exported to any OTEL-compatible collector, ensuring alignment with existing telemetry pipelines. We previously blipped Langfuse, a commonly used LLM observability platform — the OpenInference SDK can log traces into Langfuse and other OpenTelemetry-compatible observability platforms.
-
113. Valibot
Valibot is a schema validation library in TypeScript. Like other popular TypeScript validation libraries such as Zod and Ajv, it provides type inference, but its modular design sets it apart. This architecture allows bundlers to perform effective tree shaking and code splitting, including only the validation functions actually used. Valibot can reduce the bundle size by up to 95% compared to Zod in optimal scenarios. It’s an appealing choice for schema validation in environments where bundle size is critical, such as client-side validation or serverless functions.
-
114. Vercel AI SDK
Vercel AI SDK is an open-source, full-stack toolkit for building AI-powered applications and agents in the TypeScript ecosystem. It consists of two main components: AI SDK Core standardizes model-agnostic LLM calls, supporting text generation, structured object generation and tool calling; AI SDK UI simplifies front-end development with streaming, state management and real-time UI updates in React, Vue, Next.js and Svelte, similar to assistant-ui. For teams already working within the TypeScript and Next.js ecosystem, Vercel AI SDK provides a fast, seamless way to build AI applications with rich, client-side experiences.
暂缓
无法找到需要的信息？
每期技术雷达中的条目都在试图反映我们在过去六个月中的技术洞见，或许你所搜索的内容已经在前几期中出现过。由于我们有太多想要谈论的内容，有时候不得不剔除一些长期没有发生变化的条目。技术雷达来自于我们的主观经验，而非全面的市场分析，所以你可能会找不到自己最在意的技术条目。
