Zero-Shot Prompt Routing with Liquid AI's LFM 2.5-Encoder-350M

Melvin Vivas · X video post · 2026-09-27 · 2:16 · 510 views · Open on X

Topics: LLMOps, Deployment & Monitoring, AI System Design & Architecture, LLM Fundamentals · Level: intermediate

Summary

This short demo shows how Liquid AI's LFM 2.5-Encoder-350M routes prompts zero-shot. It sends simple requests to cheap or small models and complex ones to expensive models. The categories are supplied at runtime, so you can add, remove or rewrite routes without training a new classifier. The same mechanism can catch off-topic or low-value requests and send them to a cheaper model, a safe default, or a decline.

Key points

• Core idea: not every request needs the largest, most expensive model. Send complex tasks to strong models and simple ones to small or free models.
• Example device-assistant routes: a weather query or setting a timer is a simple function call, while planning and booking a trip is a complex multi-step agentic task.
• Other routing examples: send code prompts to the right programming language, route math prompts, or classify prompts by topic.
• How it works: the encoder reads the full prompt and scores it against every category in one forward pass. It runs locally and almost instantly.
• It is zero-shot: categories are given at runtime, so adding a new route (for example 'soccer agent') needs no retraining.
• Demo: 'Who won the 2026 FIFA World Cup?' first went to the complex agentic task route. After a 'soccer agent' category was added at runtime, it went to the soccer agent, while 'set a timer' still went to the simple route.
• Policy use: the same router can catch off-topic or low-value requests and send them to a cheaper model or a safe default, or decline them. This lets you write routing policy specific to your use case.

Resources mentioned

• LFM2.5-Encoder-350M · tool · huggingface.co · free
  A 350M-parameter encoder model from Liquid AI that does zero-shot prompt classification and routing against categories you supply at runtime.
  Also in: AIBackends v0.7.0: Prompt Routing with Liquid AI's LFM2.5 Encoder (Melvin Vivas on X · notes), Model Routing Fine-Tuned on Your Agent Harness Traces (Melvin Vivas on X · notes), Zero-Shot Prompt Routing with Liquid AI's LFM 2.5-Encoder-350M (Melvin Vivas on X · notes), Zero-Shot Prompt Routing by Task Complexity with LFM2.5-Encoder (Melvin Vivas on X · notes)
• Liquid AI (@liquidai) on X · website · x.com · free
  An AI company that builds efficient foundation models. The quoted post shows its PII handling working on Japanese text.
  Also in: Liquid AI LFM 2.5 Encoder: a CPU-friendly encoder model (Melvin Vivas on X · notes), Fine-tuning Liquid AI LFM2/LFM2.5 MoE models with the new Halo framework (Melvin Vivas on X · notes), Liquid AI's LFM2-Longevity models for aging-data analysis (Melvin Vivas on X · notes), Local Speech-to-Text with LFM2.5-Audio-1.5B and llama.cpp (Melvin Vivas on X · notes) and 19 more

Try this

• Try LFM 2.5-Encoder-350M as a zero-shot router that sends simple requests to cheap models and complex ones to expensive models.
• Define routing categories at runtime and add, remove or rewrite them as your use case changes, with no classifier retraining.
• Add routes that catch off-topic or low-value requests and send them to a cheaper model or a safe default, or decline them.
• A device-assistant router that sends simple function calls (weather, timers) to a small model and complex multi-step agentic tasks (trip planning and booking) to a large model.
• A topic router that sends prompts to specialist agents (for example a soccer agent) added at runtime.
• A coding or math prompt router that sends each prompt to the right language-specific or domain-specific handler.