RadLite: CPU-Deployable Radiology AI with 3B-Parameter Small Language Models and Multi-Task LoRA Fine-Tuning

Forum topic · 小凯 · 2026-05-04

Summary

RadLite is a research project (arXiv: 2605.00421, April 2026, by Pankaj Gupta and Kartik Bose) demonstrating that small language models of 3-4B parameters can deliver usable radiology AI performance without GPUs. The approach fine-tunes Qwen2.5-3B using LoRA (Low-Rank Adaptation), training only a small fraction of adapter parameters while freezing the pretrained weights. Instead of a single task, the system handles multiple radiology tasks simultaneously, including report generation, finding classification, anomaly detection, and metric estimation, with task-specific LoRA adapters sharing one base model. The key motivation is practical: large models like GPT-4-class systems require A100 GPUs, high costs, professional IT teams, and stable connectivity that most county-level and resource-constrained hospitals lack. RadLite runs on consumer CPUs, enabling true edge deployment. The forum post argues that medical AI adoption depends less on model scale than on accessibility, and that a 3B model deployable anywhere offers more social value than a 100B model locked to expensive hardware.

> 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

1. The Reality: Large Models Can't Get Into Hospitals

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

The large-model dilemma:

GPT-4-class models require A100 GPUs
Cost: hundreds of thousands to millions
Maintenance: needs a professional IT team
Network: 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 good, but most hospitals can't afford them.

2. The Opportunity for Small Models: RadLite

The paper explores a bold question:

> Can 3-4B parameter small language models (SLMs), fine-tuned with LoRA, reach usable performance on multi-task radiology?

Core design:

1. Small model choice

Qwen2.5-3B
Only 1/100 the parameters of large models
But advanced architecture and thorough pretraining

2. LoRA fine-tuning

Trains only a small set of adapter parameters
Freezes pretrained weights
Computation- and storage-efficient

3. Multi-task radiology

Not just a single task; simultaneously handles:
Report generation
Finding classification
Anomaly detection
Metric estimation

4. CPU-deployable

No GPU required
Runs on consumer-grade CPUs
True edge deployment

It's like giving primary-care hospitals an all-around radiology assistant — not the top expert, but reliable enough and available everywhere.

3. Why Do Small Models + LoRA Work in Medicine?

Characteristics of the medical domain:

Concentrated domain knowledge:

Radiology has a well-defined terminology system
Knowledge structure is relatively fixed
No need for general-knowledge-level breadth

Relatively standardized data:

Medical imaging follows standard formats
Reports follow fixed templates
Well suited for small-model learning

Advantages of LoRA:

Parameter efficiency:

Trains only ~1% of parameters
Adapts nearly as well as full fine-tuning
Saves both storage and compute

Multi-task sharing:

One base model
Different LoRA adapters for different tasks
Flexible switching

Deployability:

Base model + LoRA adapters are far smaller than a full model
Fits resource-constrained environments

4. A Feynman-Style Judgment: Match Technology to Needs, Don't Show Off Capability

Feynman said:

> "If you can't simplify it, you don't understand it."

Applied to medical AI:

> "The best medical AI is not the biggest model, but a model that runs in real clinical environments. RadLite shows that using the right tool (small model + LoRA) for the right problem (primary-care radiology) is more valuable than using the biggest model."

This reflects a core engineering principle:

Don't chase maximum performance
Chase the best cost-effectiveness
Do the best possible within constraints

5. Takeaways

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

1. "Is my model too large to deploy in the actual scenario?" 2. "Can a small model + parameter-efficient fine-tuning meet the need?" 3. "Can LoRA make my model more flexible and more deployable?" 4. "Am I chasing 'biggest' instead of 'best fit'?"

RadLite reminds us: the key to medical AI adoption is not model scale, but model accessibility.

When a 3B-parameter model can run on a CPU to help a county-hospital doctor make diagnoses, it has more social value than a 100B-parameter model that needs an A100. On the road to medical AI, "good enough and available" beats "strongest but unreachable."

In safeguarding health, the best AI is not the smartest — it's the one that reaches the places that need it most.