Ant Group open-sources LemonDecider, a 4B decision model beating Jev 1.14 at 50.7ms median latency
aigclink · x · 2026-11-09
Ant Group open-sourced LemonDecider, a decision-specialized model built on Qwen3.5-4B. Instead of generating text and parsing the answer, the model directly outputs the judgment, cutting latency sharply.
Three output types:
- Choice: pick one option — which tool to call, which workflow to enter
- Yes/No: binary judgment, e.g. sufficient evidence, request eligibility
- Score: rate retrieval relevance, content quality, etc.
A notable trick: multiple decision questions can be submitted against the same context and answered in a single forward pass when length allows, avoiding redundant context processing.
Benchmarks: 61.67 on Decision Index 0.4.1 Public, beating Jev 1.14's 60.70; median inference latency of 50.7ms across 140k test requests. The release includes 1.2M training samples and training code, covering 864 task tags including intent recognition, tool routing, evidence judgment, retrieval relevance and content safety.
More from Models
- Together AI: leaderboards are for shortlisting, your production traffic is the real benchmark — togethercompute · 2026-11-09
- Gemini quietly reads your Android messages, calls and Gmail — here's how to limit it — Med1_Ai · 2026-11-09
- Airbnb: fine-tuned small model matches frontier quality on production classifiers at ~1% of the cost — lak_luster · 2026-11-09
- Kaggle team position paper: AI competitions are the gold standard for GenAI evaluation — MeganRisdal · 2026-11-09
- Internal 32B/122B/256B checkpoints rumor debunked as AI-generated fake news — teortaxesTex · 2026-11-09
- Joseph Jacks: Liquid AI is the only US frontier lab keeping open weights ahead of China — JosephJacks_ · 2026-11-09