Post

Instead of generating text token by token, "System 1" decision models (like Jev) act as fast, deterministic decision engines. In a single forward pass, they output calibrated probabilities. Here's how I turned DiffusionGemma into a multimodal visual classifier in Java! 🧵👇

Guillaume Laforge

@glaforge.dev

The open-source foundation: Matt Mastracci (github.com/mmastrac) built djev on #DeepMind's DiffusionGemma. Using Daniel Lee's recipe (github.com/taeold), I deployed it on Cloud Run with NVIDIA Blackwell GPUs. With min-instances=0, it scales to zero idle (answering in ~140 ms!)

github.comtaeold - Overviewtaeold has 80 repositories available. Follow their code on GitHub.

Guillaume Laforge

@glaforge.dev

What about images? DiffusionGemma includes a SigLIP vision tower. I turned vision back on in Cloud Run and sent a photo of my cat: 😺 Is it a cat? (99.99%) 👁️ Eye color: Blue (99.99%) ☁️ Fluffiness: Extra fluffy All three questions evaluated concurrently in just 129 ms! 🕑

Guillaume Laforge

@glaforge.dev

Bringing it to #Java with @LangChain4j.dev: LangChain4j was adding StructuredDecisionModel for text. I proposed & implemented native multimodal support. In Java, pass ImageContent with typed Question models (Choice, Noul, Score) for clean, zero-shot visual classification!

24/9/26, 7:52 pm0 reposts1 like

Guillaume Laforge

@glaforge.dev

Full story, architecture & code: 👉 glaforge.dev/posts/2026/... Kudos to Matt, & Daniel Lee! Where could you use instant decision models instead of heavy LLMs? Let me know!

glaforge.devMultimodal Decision Models: From DiffusionGemma and Jev to LangChain4jHow I adapted Matt Mastracci's djev and Daniel Lee's Cloud Run deployment to run multimodal visual classification on Blackwell GPUs, and proposed multimodal extensions to LangChain4j's StructuredDecisionModel.