Ant open-sources LemonDecider, a 4B decision model clocking 50.7ms

LemonDecider is an open-source 4B specialized decision model from Ant, trained on Qwen3.5-4B. Instead of spitting out conversational text and relying on parsers to extract an answer, it outputs decisions directly.

What it does and how it works

The model natively handles three output types: Choice for routing between tools or workflows, Yes/No for binary checks like evidence sufficiency and compliance, and Score for assessing retrieval relevance or content quality.

It also supports batched decision-making over shared context. You can submit multiple questions against the same prompt, and as long as context length permits, the model resolves them in a single forward pass without re-encoding the background tokens.

On the Decision Index 0.4.1 Public benchmark, LemonDecider 4B scored 61.67, edging past Jev 1.14 at 60.70. Across roughly 140,000 test requests, median inference latency landed at 50.7ms. Ant open-sourced the model weights alongside training code and 1.2 million training samples covering 864 task tags like intent detection and safety filtering.

Why it matters

If you build AI agents or RAG pipelines, pinging frontier models just for tool routing or relevance scoring burns cash and adds unnecessary latency. At 4B parameters and around 50ms per run, this model makes a practical, lightweight traffic controller.

The real prize might actually be the data. Releasing 1.2 million decision-focused training samples and the training pipeline gives developers a solid blueprint to train their own specialized routing models in-house.