Antares-1B
#113 in Open-Source-SpraakmodelleCisco · 1b · 2× · tolest 24. Juli 2026
4
Momentum
Antares-1B is a openly weighted 1-billion-parameter language model developed by Cisco (Foundation AI), specifically trained for vulnerability localization in source code repositories. It is based on the IBM Granite 4.0 1B backbone and operates as a terminal agent that autonomously searches repositories using shell commands (grep, find, cat) to identify suspicious files. The model was released alongside the smaller Antares-350M on July 21, 2026 under Apache 2.0 license on Hugging Face; a larger Antares-3B model remains unpublished for now. Cisco positions Antares as a cost-effective, locally deployable alternative to large cloud AI models for security teams.
Momentum-Verloop
11.06.09.09.
Features
| Key Benchmark (%) | 0.209 File F1 on Cisco VLoc Bench (500 tasks); highest recall of any evaluated system (0.224) |
| Context Window (Tokens) | 128,000 tokens (128K) |
| License | Apache 2.0 |
| Multimodality | No – text/code-only language model (no vision/audio support) |
| Platform | Hugging Face (Transformers-compatible weights, official CLI with OpenAI-compatible inference endpoint; community GGUF available for llama.cpp) |
| Release Date | July 21, 2026 |