// open-source line
← products

CratonVM

The Java Runtime for AI Inference and Big Data at Scale

Version 0.3.0 · Java SE 8–25 · 400K+ lines of Rust · 19K+ automated tests · by Craton Software Company

Java SE 8–25
Language Coverage
400K+
Lines of Rust
19K+
Automated Tests
100% Rust
Memory Safety
AI Inference & Acceleration

Built for the AI era

Every modern AI workload lives or dies on one operation: matrix multiplication. CratonVM ships GEMM — the core computation behind every neural-network layer, every transformer attention block, every LLM inference pass — as a built-in, GPU-accelerated kernel, in both full precision (fp32) and the half precision (fp16) that makes modern inference fast.

And it doesn't stop at matmul. CratonVM's automatic GPU pipeline recognizes eligible Java compute — division-heavy math, reductions, high-resolution rendering, and more — and moves it to the GPU with nothing more than a build flag and a run flag. The same .class files. The same main(). The same results HotSpot would give you — bit for bit — just dramatically faster.

Compare that to the alternative. TornadoVM, the best-known Java GPU framework, asks you to annotate your kernels, restructure them around a TaskGraph API, and adopt new off-heap array types before a single line moves to the GPU. CratonVM asks for none of that — and still comes out ahead:

Compute kernel (16.7M+ elements)CPU (HotSpot C2)TornadoVM GPUCratonVM GPUvs HotSpotvs TornadoVM
Integer division chain2,146 ms26 ms11 ms195x faster2.4x faster
Double-precision division chain1,780 ms128 ms95 ms18.7x faster1.3x faster
128 multiply-adds/element1,300 ms27 ms8 ms163x faster3.4x faster
Dot-product reduction (int·int → long)1,172 msunimplemented12 ms98x faster
Ray tracer kernel (33.2M pixels)837 ms24.3 ms12.3 ms68x faster2.0x faster

Every number above is checksum-verified bit-for-bit against HotSpot, including the GPU results — no rounding, no approximation, no "close enough". And that dot-product row isn't a gap in our table: TornadoVM's own engine throws unimplemented on that shape. CratonVM just runs it.

High-Throughput Services

Built for Big Data — at the scale that actually matters

AI needs matrix multiplication. Big data needs throughput, uptime, and heaps that don't stall your service under load. CratonVM was engineered for both:

01

A collector built for large heaps and low pauses

CratonVM's default garbage collector is a modern, compaction-capable design sized for the memory profile real data services carry today — not a decades-old collector stretched to cover a new decade of workloads.

02

Runs the frameworks your data stack is already built on

Spring & Spring Boot, Tomcat, and Hibernate boot, run, and pass their real test suites on CratonVM today, backed by a full embedded H2 SQL database. Netty, Quarkus, and Hibernate Reactive are next, bringing the full reactive, event-driven stack online.

03

One self-contained binary

No JDK to install, no rt.jar, no bloated base image. Download it and run it — in a container, at the edge, in a pipeline, wherever your data actually lives.

04

19,000+ automated tests, checked line-for-line against HotSpot

Verified against real HotSpot behavior across 19,000+ tests, so what runs on CratonVM behaves the way your engineering team already expects Java to behave.

Clean-Slate Architecture

A runtime built different, on purpose

For twenty-five years the JVM has rested on millions of lines of C and C++. CratonVM starts over — in Rust.

01

Memory-safe by construction

The interpreter, JIT compiler, and garbage collector are written in safe Rust wherever the problem allows it, closing off entire categories of memory-corruption bugs that have shadowed native runtimes for decades.

02

A real optimizing JIT

Not a toy interpreter: tiered compilation, on-stack replacement (OSR), bounds-check elimination, AVX2 vectorization — your hot loops compiled straight to native x86-64.

03

Modern Java, fully covered

Java 8 through Java 25 — lambdas and streams, records and sealed classes, pattern matching, virtual threads, and scoped values.

04

400,000+ lines of from-scratch Rust across 22 crates

One coherent runtime, not a patchwork of decades-old C++ with a new frontend bolted on.

Ecosystem Adoption

Who's already building on it

AI/ML platform teams

Teams who want GPU acceleration for Java-based inference and data-prep pipelines without rewriting a single kernel.

Data platform engineers

Engineers running Spring Boot, Hibernate, and H2 at scale, who want a lower-pause, more predictable GC underneath them.

Performance engineers

Engineers tired of opaque, decades-old JVM internals — CratonVM is transparent, hackable, and built to be understood.

Security-conscious organizations

Organizations who want their runtime's memory-safety story to be a design decision, not a patch cycle.

Get started in sixty seconds

bash
cargo build --release -p cratonvm-cli
./target/release/cratonvm -cp . HelloWorld

CratonVM: the Java runtime built for what's next. Built in Rust, by Craton Software Company.

CratonVM is under active, fast-moving development. Java and OpenJDK are trademarks of Oracle and/or its affiliates. NVIDIA and CUDA are trademarks of NVIDIA Corporation. TornadoVM is a project of the Beehive Lab, University of Manchester. These references are used for compatibility and comparison identification only.

Ready to secure
the future?

Request Expert Briefing