340M Model Beats 54B Decoder

🔥 What happened Fastino Labs dropped GLiNER2.5-Decide: a 340M-parameter decision model under Apache 2.0 that takes text plus a typed schema and returns structured answers with probability distributions and confidence scores. Runs on CPU, GPU, or air-gapped. 💡 Why it matters It beats a 54B-class decoder (60.1% vs. 57.5%) on intent routing while being 160x smaller. p50 latency: 167 ms on CPU, 38 ms on V100. Joint decoding stops contradictory outputs like "safe" and "prompt_injection" firing at once. ⚡ Our take If you're still burning a 70B LLM for routing and guardrails, you either have too much money or too little latency budget. This belongs in every agent pipeline.