
Using @GraalVM native image to AOT compile Java base and javac is still giving another 3-4x better numbers. And this without the need for a training run. github.com/thomaswue/crema-s…

"AI in a jar" @ Quixotic AI

Using @GraalVM native image to AOT compile Java base and javac is still giving another 3-4x better numbers. And this without the need for a training run. github.com/thomaswue/crema-s…

The next generation of @GraalVM is shaping up nicely: fast startup, low CPU and memory usage, with no build step or project changes required! 🚀 Here’s a simple benchmark using a single-file program:

Spring Boot shops, banks, insurers, payroll systems, all the backends that quietly run the world: millions of JVM apps can now add a 24B-A2B reasoning model as a plain Maven dependency. Every Liquid AI model is supported. No infra. No ops. No Python. 1009 tok/s prefill (~2.3x llama.cpp) 47.8 tok/s decode (parity) qxotic.ai
Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest 𝙻𝙵𝙼𝟸.𝟻 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses. > 24B total parameters, 2B active per token > Context expanded to 131K tokens > A 128K-token vocabulary > 33T pretraining tokens > It is a reasoning model > Very easy to fine-tune and customize > Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks. > Open base and post-trained weights on @huggingface 🤗 1/

The End of Programming? Video of my talk yesterday at #scaladays. youtube.com/watch?v=jlJyuJhD…

The first #Devoxx talk I watched offline! 🤩 Super interesting presentation and amazing work! ☕️👏 #Congrats This could be a great fit for real-time transcripts at #Devoxx Belgium 2027.... Also excited to explore how we could integrate it with @DevoxxGenie 🚀

The video is out! My #Devoxx talk, "Quixotic: Pure Java Inference at Native Speed" is now available: youtu.be/S7x4DxwRnk4



AI on the JVM: Quixotic? How fast can an LLM written in pure Java really be? Faster than C++? All the answers today, 11:50, Room 4. #Devoxx m.devoxx.com/events/dvbe26/t…

It's a pity JVMCI was removed from OpenJDK now that Graal CE JIT gets faster than C2. Graal CE 25.4.4 Native Image performance improved significantly. The last missing feature is probably PGO. Shenandoah for NI is coming along nicely too #graalvm @graalvm github.com/simonis/Renaissan…

M_AI_cronaut?
Think we should consider renaming Micronaut Framework to Micronaut Intelligence Cache AI edition 😎

Think we should consider renaming Micronaut Framework to Micronaut Intelligence Cache AI edition 😎

I can't be the only one thinking @maxandersen deserves a Hall Of Fame award for single handedly making Java CLIs usable/viable with @jbangdev. It's such a massive lever and key to the story.

"Write once, run anywhere!" Not "Write once, bridge to Python!" qxotic.ai/#examples

AI in Java, at native speed, competitive with native engines on CPUs. qxotic.ai/#benchmarks
AI in Java... quixotic, until now! End-to-end AI stack for the JVM, built from first principles: Tokenizers, quantized matrix multiplication, model format readers, inference engine, vision, speech: all in Java. No external services, no wrappers, no Python, no ONNX. Will be presented at Devoxx Belgium, next week: "Quixotic: Pure Java Inference at Native Speed" m.devoxx.com/events/dvbe26/t…

Meet Tok'n'Roll 🎸 the tokenization substrate of qxotic.ai Token-perfect. Pure Java. Zero dependencies.