GPULlama3.java 1.0.0 Reaches General Availability
GPULlama3.java, a GPU-accelerated Llama 3 inference project built on TornadoVM, has shipped its 1.0.0 GA release. The headline addition is support for the TornadoVM CUDA backend with accelerated batch prefill powered by Tensor Core MMA (matrix multiply-accumulate) instructions, which speeds up the prompt-processing phase on NVIDIA hardware significantly compared to earlier releases that leaned on OpenCL.
The release also adds an OpenAI-compatible HTTP server, so existing tooling built against the OpenAI chat completions API can point at a local GPULlama3.java instance without modification. A new RunMetrics class exposes structured performance data — tokens per second, prefill time, and decode latency — making it easier to benchmark different GPU backends from the same harness.
For Java developers exploring local LLM inference without leaving the JVM, GPULlama3.java is one of the few projects offering native GPU acceleration through TornadoVM rather than shelling out to a separate Python or C++ runtime. The 1.0.0 label signals the maintainers consider the CUDA and OpenCL backends production-stable.
Read more — InfoQ
Maven 4.0.0 RC6 Refines Multi-Module Artifact Handling
The Apache Maven team shipped the sixth release candidate of Maven 4.0.0. The most notable change lets projects accept Java module names as an attached artifactId, even when that name differs from the project's own artifactId — a fix aimed at multi-module builds that need to publish multiple artifacts with distinct Java module identities from a single project.
RC6 also extracts maven-executor into its own standalone project, separating execution logic from the core build engine. This should make embedded and forked execution modes easier to maintain and test independently of the rest of Maven's codebase, and it opens the door for other tools to reuse the executor without pulling in the full Maven distribution.
With six release candidates now shipped, Maven 4.0.0 appears to be converging on a final release. Teams still running Maven 3.x should start test-migrating build configurations against the RC builds now, since Maven 4 introduces changes to the POM model and plugin execution semantics that are easier to catch early than after a GA cutover.
Read more — InfoQ
Jakarta Agentic AI 1.0.0 Ships First Milestone
The Jakarta EE Working Group released the first milestone of Jakarta Agentic AI 1.0.0, a new specification aimed at standardizing how Jakarta EE applications build and integrate AI agents. This milestone focuses on test infrastructure: the TCK (Technology Compatibility Kit) now moves beyond simple signature checks to validate actual runtime behavior.
A notable addition is the @RequiresEngine and @RequiresNoEngine pair of annotations, which replace JUnit's generic @Disabled annotation for tests that depend on (or must run without) a specific agent execution engine being present. This gives TCK implementers finer-grained control over conditional test execution than a blanket skip would allow.
Alongside the milestone, the Eclipse Starter for Jakarta EE was updated to support Jakarta EE 11, giving developers a way to scaffold projects that can immediately pull in the new agentic AI APIs once they stabilize. This is an early milestone rather than a finished spec, but it marks Jakarta EE's first formal step toward standardizing agent-building patterns that have so far been the domain of framework-specific libraries like LangChain4j and Spring AI.
Read more — InfoQ
Framework Point Releases: Micronaut 5.1.0, Quarkus 3.38.0, and JobRunr 8.8.0
Three Java framework projects shipped point releases in late July. Micronaut Framework 5.1.0, built on Micronaut Core 5.1.10, adds a new @Property annotation for sequenced-collection injection and support for OpenDI, an Eclipse-compatible implementation of CDI Lite — giving Micronaut users a lighter-weight dependency injection option that aligns with Jakarta CDI conventions.
Quarkus 3.38.0 introduces weight-based memory eviction for Hibernate second-level cache regions, configurable via either Caffeine or JCache, letting teams bound cache memory usage by weight rather than just entry count. The release also adds an HTTP Problem extension implementing RFC 7807 for standardized machine-readable error responses across REST endpoints.
JobRunr 8.8.0, the background job processing library, improved logging around StorageException handling for easier debugging of persistence failures, added Kotlin 2.4 support, and removed the hard dependency on quarkus-smallrye-health for Quarkus users who don't need health-check integration.
Read more — InfoQ
Links & Sources
You May Also Enjoy
