TornadoVM 4.0 GA: Apple Silicon Metal, CUDA Graphs, JDK 25/21
TornadoVM 4.0.0 reached general availability on April 2, 2026, delivering heterogeneous GPU acceleration for Java applications with two headline additions. The first is a new Apple Silicon backend powered by the Metal API, meaning Java code can now run on the GPU of any M-series Mac without requiring OpenCL or third-party shims. The second is an enhanced PTX backend with SIMD shuffle and reduction intrinsics, improving throughput for numeric workloads on Nvidia hardware. A new withCUDAGraph() method on TornadoExecutionPlan allows the runtime to capture a sequence of CUDA operations into a graph and replay it with minimal driver overhead on subsequent iterations.
The release supports both JDK 25 and JDK 21, making it accessible to teams on the current LTS and the latest feature release. TornadoVM's programming model remains annotation-driven: developers annotate hot loops with @Parallel and the runtime transparently compiles and dispatches them to available accelerators — CPUs, Nvidia GPUs, AMD GPUs, Intel integrated graphics, and now Apple Silicon — selecting the fastest device at runtime. The framework includes live profiling utilities, dynamic task migration between devices, and a batch-processing facility for datasets that exceed GPU memory.
TornadoVM 4.0 also ships GPULlama3.java, a Java-native Llama 3 inference implementation that compiles directly to GPU kernels via TornadoVM, eliminating the JNI bridge to native inference libraries that most Java LLM integrations depend on today.
Read more — InfoQ
Google ADK for Java 1.0 GA: Stable AI Agent Framework with A2A Protocol
Google released version 1.0.0 of the Agent Development Kit (ADK) for Java on March 30, 2026, marking the first stable release of its open-source framework for building production-ready AI agents in Java. The ADK provides code-first, framework-agnostic primitives for agent construction and has grown from a Python-only tool into a multi-language ecosystem covering Python, Java, Go, and TypeScript.
The 1.0 release adds several production-critical features. Human-in-the-Loop (HITL) workflows are built around a ToolConfirmation mechanism: when a tool requires human approval before execution, the agent pauses, surfaces the pending action to the user, and resumes on confirmation — enabling safe deployment of agents that modify real resources. The release also introduces native support for the Agent-to-Agent (A2A) protocol via the official A2A Java SDK client, so ADK-built agents can discover and communicate with agents on other frameworks using the standardised JSON/HTTP handshake. Session management gains three concrete service implementations — InMemorySessionService for local development, VertexAiSessionService backed by Cloud's Vertex AI Session API, and FirestoreSessionService for persistent storage.
The InMemoryArtifactService is now integrated into AgentExecutorProducer, providing a default artifact store for agents that generate files or structured outputs during a run. For models that do not natively support using output_schema and tools simultaneously, ADK 1.0 adds a compatibility layer that emulates both features in a single call, removing a common source of capability gaps when targeting non-Gemini backends.
Read more — Google Developers Blog
Grails 7.1.0 RC1 and Gradle 9.5.0 RC1
The first release candidates of Grails 7.1.0 and Gradle 9.5.0 both shipped in the March 30 Java news cycle, offering early adopters a preview of upcoming stable releases.
Grails 7.1.0 RC1 centralises Groovy invokedynamic configuration inside the Gradle Plugin, ending the fragmented approach where teams configured it per-module. The @Service annotation now automatically inherits the datasource mapping from its associated domain class rather than requiring an explicit declaration, reducing boilerplate in service definitions for multi-datasource applications.
Gradle 9.5.0 RC1 focuses on diagnostics and collection safety. Task failure output now includes provenance information — the plugin or build script that registered the failing task — making it easier to attribute broken builds to their origin. A new disallowChanges() method on DomainObjectCollection allows build authors to lock collections at configuration time, catching accidental late mutations that cause hard-to-reproduce build failures.
Read more — InfoQ
Java Annotated Monthly — April 2026
JetBrains' Java Annotated Monthly for April 2026, curated by Marit van Dijk (Java Champion, JetBrains Developer Advocate), centres on the community's continued digest of Java 26, which shipped March 17. The edition rounds up Kotlin 2.3.20, released alongside continued ecosystem updates, and highlights the debut of Koog, JetBrains' AI agent framework for Java, which brings agent orchestration primitives designed around JVM idioms rather than porting Python patterns.
Notable articles include an analysis of JDK Flight Recorder combined with AI-assisted anomaly detection for intelligent JVM monitoring, and architectural guidance on "Moving beyond Strings in Spring Data." The event calendar features Spring I/O in Barcelona (April 13–15), Java Day Istanbul (April 17–18), JCON EUROPE in Cologne (April 20–23), and Devoxx France and Greece (April 22–25). The edition also marks 25 years of IntelliJ IDEA, celebrated with a documentary release and a JavaOne keynote.
Apache Tomcat maintenance releases — versions 11.0.21, 10.1.54, and 9.0.117 — corrected a non-blocking flush bug in the NIO/TLS stack where responses could remain unwritten until connection close, and improved EncryptInterceptor and HTTP/2 error handling. Apache Log4j 2.25.4 fixed RFC5424Layout divergence between documentation and implementation.
Read more — The IntelliJ IDEA Blog
Links & Sources
You May Also Enjoy
