RadLite: Multi-Task LoRA Fine-Tuning Brings Radiology AI to 3B Models on CPUs

Forum topic · 小凯 · 2026-05-04

Summary

RadLite is a paper (arXiv: 2605.00421) by Pankaj Gupta and Kartik Bose exploring whether small language models of 3–4B parameters can handle multiple radiology tasks after LoRA fine-tuning while running entirely on consumer CPUs. The post argues that frontier-scale models like GPT-4 require A100 GPUs, large budgets, professional IT teams, and stable high-speed networks—resources unavailable to most county-level hospitals. RadLite's approach combines Qwen2.5-3B as the base model, parameter-efficient LoRA adapters (training only ~1% of parameters) that are swapped per task, and multi-task coverage including report generation, findings classification, abnormality detection, and metric estimation. The resulting system deploys on CPU without a GPU, enabling true edge deployment in resource-constrained medical settings. The author frames this as an engineering lesson: medical AI succeeds through accessibility and fit-for-purpose design rather than raw scale—a 3B model running on a CPU can deliver more social value than a 100B model locked behind expensive hardware. Key takeaways for builders: question whether your model is too large to deploy, whether LoRA-based parameter-efficient fine-tuning suffices, and whether optimizing for 'best fit under constraints' beats chasing maximum performance.

Overview

Paper: RadLite: Multi-Task LoRA Fine-Tuning of Small Language Models for CPU-Deployable Radiology AI Authors: Pankaj Gupta, Kartik Bose arXiv: 2605.00421 | 2026-04-29

This is a translation/summary of a Chinese forum post discussing the paper.

The Problem: Large Models Can't Reach Most Hospitals

Imagine a county-level hospital wanting to deploy AI-assisted radiology diagnosis:

The dilemma of large models:

GPT-4-class models require A100 GPUs
Cost: hundreds of thousands to millions
Maintenance: needs a professional IT team
Networking: requires stable, high-speed connectivity

The reality:

County hospitals have limited budgets
No GPU servers
Insufficient IT staff
Unstable networks

Conclusion: large models may be powerful, but most hospitals can't afford them.

The Opportunity: RadLite

The paper asks a bold question:

> Can a 3–4B parameter small language model (SLM), fine-tuned with LoRA, reach usable performance on multi-task radiology workloads?

Core design:

1. Small base model — Qwen2.5-3B: only ~1/100 the parameters of frontier models, but with a modern architecture and thorough pre-training. 2. LoRA fine-tuning — trains only a small set of adapter parameters while freezing pre-trained weights, saving compute and storage. 3. Multi-task radiology — not a single-task system; it simultaneously handles:

Report generation
Findings classification
Abnormality detection
Metric estimation

CPU-deployable

Think of it as an all-in-one radiology assistant for grassroots hospitals—not a top specialist, but reliable and available everywhere.

Why Small Model + LoRA Works in Medicine

Characteristics of the medical domain:

Concentrated domain knowledge: radiology has a well-defined terminology system and relatively fixed knowledge structure; it doesn't require encyclopedic general knowledge.
Relatively standardized data: medical imaging has standard formats and reports follow fixed templates—well suited for small-model learning.

LoRA advantages:

Parameter efficiency: trains only ~1% of parameters, yet adapts close to full fine-tuning; saves storage and compute.
Multi-task sharing: one base model, different LoRA adapters per task, flexibly swappable.
Deployability: base model + adapters is far smaller than the full model—ideal for resource-constrained environments.

The Engineering Judgment: Match Technology to Need

Echoing Feynman's "If you can't simplify it, you don't understand it":

> The best medical AI is not the largest model, but the model that can run in real clinical environments. RadLite shows that using the right tools (small model + LoRA) to solve the right problem (grassroots radiology) is more valuable than using the biggest model.

Core engineering principles:

Don't chase maximum performance; chase the best cost-effectiveness
Do the best possible under constraints

Takeaways for Builders

If you're building medical AI or edge AI systems, ask yourself:

1. Is my model too large to deploy in real-world scenarios? 2. Can a small model + parameter-efficient fine-tuning meet the need? 3. Would LoRA make my model more flexible and deployable? 4. Am I pursuing "largest" instead of "most suitable"?

The key lesson: what matters in medical AI is not model scale but model accessibility. When a 3B model can help doctors at county hospitals diagnose on a CPU, it delivers more social value than a 100B model that requires an A100. In medical AI, "good enough and reachable" beats "strongest but unreachable." The best AI in healthcare is not the smartest—it's the one that reaches the places that need it.