TechPrufer
Get started
Hugging Face
Hugging Face logoHugging Face · Feature trail

Agent Model Routing Optimization

4 posts · Mar 2026 – Oct 2026

What this trail covers

Hugging Face now supports advanced agent model routing, moving beyond simple classification to a systems optimization problem. This capability allows for dynamic routing based on a multi-dimensional optimization of cost, quality, latency, compliance, and reliability, rather than just sticker price or perceived task difficulty. The system aims to make model selection an intrinsic part of the agent harness, optimizing the entire system's operating point. This post introduces Ecom-RLVE, an adaptive decision model family (d1-3B and d1-omni-600M) for edge deployment, supporting multimodal inputs (text, image, audio) and achieving high decision quality with low latency on edge devices. The models are built on Liquid Foundation Models (LFMs) and are designed for fast, structured decisions.

All Hugging Face trails
  1. Open-sourcing AstaBrief, the fast report-generation model in AstaThis post introduces AstaBrief 8B, an open-weights model specifically trained for scientific report generation. It details the training process, including collecting SFT data from real user queries and creating DPO pairs using multiple judge models. The post highlights the model's speed improvements by generating reports in a single pass and discusses the importance of grounding in evidence and citation. It also mentions the release of an example workflow for local report generation from PDFs.
  2. The Agent Said It Was Done. The Database Disagreed.This post introduces the ThinkingBox benchmark and grading system for evaluating AI agent reliability and consistency. It highlights the critical gap between an agent's reported success (tool calls, final responses) and the actual terminal backend state and side effects it leaves behind. The benchmark runs 507 stateful business workflows 20 times each, measuring pass@1, pass@20, and observed 20/20. Key findings include that many agents fail executable checks despite clean tool calls and no reported errors, with significant discrepancies in field values and unintended side effects. The post details the cost of consistency and provides instructions on how to run the benchmark using OpenEnv and Typesense, and discusses the implications for agent development and deployment.
  3. Multimodal open d1 decision models for the edgeThis post introduces the d1 decision model family (d1-3B and d1-omni-600M) for edge deployment. d1-3B is built on LFM2.5-VL-3B (decoder-only, text+image input) and d1-omni-600M is built on LFM2.5-Encoder-350M (bidirectional encoder, text+image or text+audio input). Both models are designed for single forward pass decision making, not token generation. Benchmarks show d1-3B achieving high scores on various decision tasks and outperforming larger models. Performance metrics on NVIDIA edge devices (Jetson AGX Thor, Orin, Nano) and GPUs (RTX 4090, AMD MI325X) demonstrate low latency for single and multiple questions, as well as image processing. The post provides Python code examples for loading and using the models, including batching capabilities.
  4. Model Routing Is Simple. Until It Isn’t.This post introduces a novel approach to agent model routing, reframing it from a classification problem to a systems optimization problem. It highlights that effective routing must consider factors beyond model pricing, including caching behavior, workload patterns, infrastructure state, and compliance constraints. The authors detail their experience building a router that optimizes across cost, quality, and latency simultaneously, demonstrating its effectiveness on the AppWorld Test Challenge with a CodeAct agent. The optimization-based router provides a range of operating points, allowing users to prioritize cost, latency, or accuracy, and is designed to be lightweight to avoid becoming a bottleneck.
  5. Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational AgentsThis post introduces Ecom-RLVE, an adaptive verifiable environment for e-commerce conversational agents. It details how reinforcement learning is used to optimize agent behavior for verifiable outcomes in e-commerce scenarios, focusing on adaptive learning and verifiable environments to improve agent performance and trustworthiness. The system uses a novel reward function that incorporates verifiability metrics to guide the agent's learning process, ensuring that its actions lead to demonstrable and trustworthy results.