Table of Contents
AWS Comprehend is the managed NLP service: sentiment analysis, entity recognition, key phrases, custom classification. By 2026, self-hosted small LLMs (Phi-3 Mini, Llama 3.2 3B) handle the same tasks at much lower cost with comparable quality.
Comprehend wins for: AWS-aligned shops, zero-ops, integrated with other AWS services. Self-hosted Phi-3 Mini wins for: dramatically lower cost (~£0.0001 per request vs £0.0001/100 chars), full control, custom domain fine-tuning. For volume above ~100K requests/month, self-hosted dominates economically.
Comparison
| Aspect | AWS Comprehend | Self-hosted Phi-3 Mini |
|---|---|---|
| Cost | ~£0.0001 per 100 chars | ~£0.0001 per request (whole doc) |
| Quality on standard tasks | Strong | Comparable |
| Custom domain | Custom classifier (separate cost) | Free fine-tune |
| Languages | Many | Many (Qwen for Asian; Llama for European) |
| Setup | Trivial | ~1 hour |
| Best for | AWS shops, low volume | High volume, custom domain |
When each
- AWS Comprehend: AWS-native shops, low volume (<100K requests/month), zero-ops priority
- Self-hosted: high volume, custom domain quality matters, residency requirement, cost-anchored
Verdict
For NLP workloads above modest volume, self-hosted small LLMs (Phi-3 Mini class) with structured-output JSON have replaced AWS Comprehend in most production deployments. The cost saving is 10-50× at scale; quality is comparable on standard tasks; custom domain fine-tuning is free instead of paid extra. Comprehend remains right for AWS-aligned, low-volume use.
Bottom line
Phi-3 Mini replaces Comprehend at scale. See Phi-3 use case.
