Cisco Unveils Lightweight SLM "Antares" Specialized for Vulnerability Detection

Cisco Systems announced on July 21 the launch of its small language model (SLM) family, “Antares,” designed specifically for detecting software vulnerabilities. The company aims to identify known bugs within existing codebases with high accuracy and speed, without relying on large frontier models.

The release includes two models: “Antares-350M” (350 million parameters) and “Antares-1B” (1 billion parameters), both of which are available as open-weight models on Hugging Face. However, access is restricted and not freely available for public download. DJ Sampath, Cisco’s Senior Vice President and General Manager of the AI Software Division, told The Register, “We have implemented gates to ensure appropriate access permissions.”

Reasons for Access Restrictions

Given the high risk of misuse of vulnerability detection models, Cisco has adopted a stringent access management policy. Currently, the models are being gradually made available to academic institutions, non-profit organizations, public sector security teams, and small to medium-sized businesses. Furthermore, both models are designed for local execution, requiring access keys to scan the target source code. This eliminates the need to send code to an AI provider’s server, as is common with cloud-based LLMs, enabling secure analysis even in highly confidential environments.

The name “Antares” is derived from the red supergiant star Antares. Cisco’s Vice President and Chief AI Scientist, Amin Karbasi, explained the analogy: “Antares is about 1,000 times larger than the Sun but appears small in the sky. Similarly, while the impact of vulnerabilities can span an entire codebase, their traces often exist in just a few lines within a single file among millions.”

Outperforming Large Models in Benchmarks

Cisco has developed its own benchmarking tests to evaluate the performance of the Antares family. These tests measure how efficiently AI models can identify security flaws within codebases. According to the company, Antares-1B outperformed Google’s Gemini 3 Pro and demonstrated similar capabilities to Z.ai’s GLM-5.2. Moreover, an unpublished Antares-3B (3 billion parameters) reportedly exceeded the performance of both GLM-5.2 and OpenAI’s GPT-5.5.

The models also demonstrate a significant speed advantage. According to Karbasi, Antares can scan 500 repositories in approximately 15 minutes, whereas frontier models take up to five hours. This efficiency translates into substantial cost savings.

As for the Antares-3B model, it will not be made publicly available. Instead, it will be responsibly distributed to select communities. Karbasi stated, “We are fully gating access to ensure that only necessary organizations can utilize it.”

Editorial Opinion

In the short term, the practicality of SLMs in the security domain has been reaffirmed. By integrating Antares with its networking and security products, Cisco could significantly streamline vulnerability detection workflows. The local execution model, which avoids external code exposure, is expected to see particularly high demand in regulated industries like finance and government. However, it is important to note that the models are specialized for detecting known bugs.

In the long term, open-weight SLMs tailored to specific domains might increasingly replace general-purpose frontier models. With advantages in speed, cost, and confidentiality, similar models could emerge in areas beyond security, such as code quality and compliance checks. That said, the dilemma of imposing access restrictions remains. While the open-weight but restricted-access approach balances transparency and safety, it will likely continue to face scrutiny.

As an editorial team, we will be closely watching how Antares is adopted in real-world scenarios and how other security vendors respond with their SLM offerings.

References

• “Cisco’s open-weight bug busters take on Google and OpenAI”, by Jessica Lyons — The Register, 2026-07-21T21:17:20.000Z (ARR)
• Source URL: https://www.theregister.com/security/2026/07/21/ciscos-open-weight-bug-busters-take-on-google-and-openai/5275817