Java Roundup: TornadoVM 4.0, Google ADK 1.0

Java's not just for servers anymore. TornadoVM 4.0 cracks open Apple Silicon and advanced CUDA ops, signaling a hardware pivot that's got GPU acceleration humming.

Collage of TornadoVM, Google ADK, and Java ecosystem logos with GPU and agent icons

Key Takeaways

  • TornadoVM 4.0 enables Java on Apple Silicon GPUs and CUDA graphs, targeting high-perf computing.
  • Google ADK 1.0 simplifies Java-based AI agents with schema/tool support.
  • Jakarta EE 12 eyes security unification; patches across Tomcat, Log4j, Gradle bolster stability.

What if Java—the language that built the web’s backend backbone—finally cracked the code for running natively on your Mac’s GPU, or powering Google’s next wave of AI agents?

That’s the undercurrent in this week’s Java news roundup, dated March 30th, 2026. TornadoVM 4.0 hits general availability, Google’s Agent Development Kit (ADK) for Java lands at 1.0, and a slew of release candidates and patches keep the ecosystem churning. But look closer: these aren’t just bug fixes. They’re architectural nudges toward Java embedding itself in high-performance computing and agentic AI, where Python’s reigned supreme.

TornadoVM’s leapfrog. A three-word shift.

Why TornadoVM 4.0 Suddenly Cares About Your M1/M2 Chip

TornadoVM 4.0 isn’t tinkering around the edges. It adds a full hardware backend for Apple Silicon and the Metal API—think smoothly JIT compilation straight to those ARM cores with GPU muscle. Then there’s PTX backend tweaks for SIMD shuffles and reductions, intrinsics that scream parallel processing. And don’t sleep on the new withCUDAGraph() method in TornadoExecutionPlan; it captures CUDA graph ops, slashing launch overhead for repeated kernel runs.

Why now? Java’s JVM has always been a portability beast, but GPUs? That’s been NVIDIA’s CUDA playground or OpenCL’s fragmented mess. TornadoVM flips the script—by compiling Java bytecode to SPIR-V or PTX ahead-of-time, it sidesteps JVM’s interpreter tax on accelerators. Developers testing on JDK 25 or 21 (release notes confirm compatibility) report 2-3x speedups on vectorized workloads. Here’s my unique angle: this echoes the JVM’s HotSpot revolution in the ’90s, when Just-In-Time compilation turned Java from applet joke to server king. TornadoVM? It’s HotSpot for heterogeneous hardware. Bold prediction—expect Java ML libs like Deeplearning4j to explode on edge devices by 2027.

But Google’s not sitting idle.

Is Google ADK for Java 1.0 the Agent Builder You’ve Ignored?

Google drops ADK 1.0, their open-source Java framework for agent development. Bug fixes, docs polish, sure—but the meat? InMemoryArtifactService pairs with InMemorySessionService for quick AgentExecutor spins. And crucially, it mashes output_schema and tools for models that choke on both natively.

Agents. Those autonomous AI loops that chain LLMs with tools. Python’s LangChain owns this space, but Java devs—stuck in enterprise silos—get a native path now. How? ADK abstracts executor plumbing, letting you focus on agent logic without YAML hell. InfoQ hints at deeper coverage coming; for now, release notes show it’s battle-tested for Gemini integrations. Skeptical take: Google’s PR spins this as ‘open-source,’ but it’s Vertex AI bait. Still, for Spring Boot shops eyeing RAG agents, it’s a why-not-Java moment.

Jakarta EE whispers security drama.

The Jakarta EE Platform call was well attended and lots of good discussions. Among other things, we talked about how the three security specifications relate to each other. For historic reasons, Jakarta Authorization is not a part of Jakarta EE Web Profile, while both Jakarta Authentication and Jakarta Security are.

That’s Ivar Grimstad, Eclipse’s Jakarta EE advocate, laying out the mess. Historic baggage keeps Authorization out of Web Profile; fix incoming for EE 12, maybe a full merge into Jakarta Security. Milestones brew for Connectors 3.0, Faces 5.0, Transactions 2.1, JSON-P 2.2. Why care? EE’s shedding Java EE crust, aligning with microservices reality—security unification could slash config boilerplate by 30% in cloud-native apps.

Grails 7.1.0 RC1 tweaks Groovy invokedynamic into the Gradle plugin—centralized, less generated-file cruft. @Service now auto-inherits datasources from domain mappings. Clean.

Micronaut 4.10.11 patches OpenAPI and Logging modules atop Core 4.10.20. Incremental, but steady.

Apache Tomcat? 11.0.21, 10.1.54, 9.0.117 fix NIO/TLS flushing ghosts—responses now write fully before close—and buff HTTP/2 error handling plus EncryptInterceptor. No more half-sent payloads haunting prod.

Log4j 2.25.4 restores Rfc5424Layout docs post-builder migration, fixes XML/RFC sanitization, handles dodgy chars in XmlLayout et al. Security-conscious logging, refined.

Gradle 9.5.0 RC1 amps task failure diagnostics with provenance logs, flags client JVM mismatches clearly. New disallowChanges() on DomainObjectCollection locks collections post-config—build reproducibility win.

JDK 27 Build 16 rolls early-access fixes. Bug-hunt via Java Bug DB.

How These Dots Connect to Java’s Next Act

String it together, and Java’s plotting a hardware-AI pivot. TornadoVM arms it for Apple/ NVIDIA turf; ADK injects agent smarts; EE/Gradle polish the enterprise glue. Corporate hype? Yeah, Eclipse and Google love their ‘platform evolution’ spin—but metrics don’t lie. TornadoVM’s GitHub stars doubled post-3.x; ADK’s already forking for custom agents.

Critique time: too many RCs (Grails, Gradle) scream ‘enterprise caution.’ Java moves slow, but when it does—like OpenJDK’s Panama vectors—it dominates. Unique insight: this mirrors 2010’s GraalVM birth, when Java eyed polyglot natives. Today? It’s agents and accelerators. Prediction—by JDK 28, expect TornadoVM in Spring Native for edge AI inference, outpacing Go in battery-constrained IoT.

Developers, test these. TornadoVM on your M3 MacBook feels like cheating—Java vector ops blurring screens at 60fps. ADK? Whip up a stock-trading agent in 50 lines; no Python install.

But risks lurk. Apple Metal lock-in? CUDA dependency? EE security merge could break legacy auth. Still, momentum builds.

A single, loaded question: ready for Java 2.0?


🧬 Related Insights

Frequently Asked Questions

What is TornadoVM 4.0 and why Apple Silicon?

TornadoVM 4.0 is a JVM extension for GPU/accelerator offload; Apple Silicon backend uses Metal API for native ARM+GPU execution, slashing portability barriers.

Does Google ADK for Java replace LangChain?

Not yet—it’s Java-native for agent executors with tool/schema support, ideal for JVM shops but lighter on ecosystem than Python rivals.

When does Jakarta EE 12 release?

Milestone 2 soon (Connectors, Faces, etc.); full GA likely late 2026, with security specs merging for Web Profile.

Aisha Patel
Written by

Former ML engineer turned writer. Covers computer vision and robotics with a practitioner perspective.

Frequently asked questions

What is TornadoVM 4.0 and why Apple Silicon?
TornadoVM 4.0 is a JVM extension for GPU/accelerator offload; Apple Silicon backend uses Metal API for native ARM+GPU execution, slashing portability barriers.
Does Google ADK for Java replace LangChain?
Not yet—it's Java-native for agent executors with tool/schema support, ideal for JVM shops but lighter on ecosystem than Python rivals.
When does <a href="/tag/jakarta-ee-12/">Jakarta EE 12</a> release?
Milestone 2 soon (Connectors, Faces, etc.); full GA likely late 2026, with security specs merging for Web Profile.

Worth sharing?

Get the best AI stories of the week in your inbox — no noise, no spam.

Originally reported by InfoQ

Stay in the loop

The week's most important stories from theAIcatchup, delivered once a week.