# Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B

Published: 2026-07-28 · Updated: 2026-07-28 · Category: Model Release

Source page: https://www.sipply.dev/story/semalith-v1-4-a-calibrated-184m-safety-classifier-achieving-f8ccdc7a · 中文: https://www.sipply.dev/zh/story/semalith-v1-4-a-calibrated-184m-safety-classifier-achieving-f8ccdc7a

## Summary

The recent release of Semalith v1.4 marks a significant advancement in safety classification technology, particularly for the financial services sector. This model, with only 184 million parameters, has achieved state-of-the-art performance in prompt-injection detection, outperforming its competitor, Llama-Guard-3-8B, which has 44 times more parameters. Semalith v1.4's ability to manage multiple safety classifications in a single inference pass highlights its efficiency and effectiveness in real-world applications.

## Key takeaways

- Semalith v1.4 is a highly efficient safety classifier with only 184M parameters.
- It outperforms Llama-Guard-3-8B in prompt-injection detection.
- The model is particularly tailored for the financial services industry.

## Why it matters

The release of Semalith v1.4 signifies a leap forward in the development of lightweight yet powerful AI models for safety classification, which is crucial for sectors like finance that require robust security measures.

- **Product managers:** For product managers, Semalith v1.4 offers a competitive edge in integrating advanced safety features into financial products without the overhead of larger models, potentially reducing costs and improving performance.
- **Developers:** Developers can leverage Semalith v1.4's efficient architecture to implement prompt-injection detection in applications with limited computational resources, enhancing security without compromising on speed.
- **Students and job seekers:** Students and job-seekers should note that advancements like Semalith v1.4 highlight the growing demand for expertise in AI safety and security, making it a valuable area of focus for future careers.

## Sources

- [Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B](https://arxiv.org/abs/2607.22545) (arXiv cs.LG, 2026-07-28)

## 中文

### Semalith v1.4：一款经过校准的184M安全分类器，以44倍于Llama-Guard-3-8B的参数实现最先进的提示注入检测

Semalith v1.4的发布标志着安全分类技术的重大进步，特别是在金融服务行业。该模型仅有1.84亿个参数，在提示注入检测方面表现出色，超越了竞争对手Llama-Guard-3-8B，其参数量是其44倍。Semalith v1.4在单次推理中管理多个安全分类的能力突显了其在实际应用中的效率和有效性。

**要点：**

- Semalith v1.4是一款高效的安全分类器，仅有1.84亿个参数。
- 它在提示注入检测方面超越了Llama-Guard-3-8B。
- 该模型特别针对金融服务行业量身定制。

**为什么重要：** Semalith v1.4的发布标志着轻量级且强大的安全分类AI模型发展的飞跃，这对需要强大安全措施的金融等行业至关重要。

---
Sipply: Sip AI News Simply. https://www.sipply.dev