---
license: apache-2.0
language:
  - en
library_name: gguf
pipeline_tag: text-generation
tags:
  - security
  - vulnerability-detection
  - agentic
  - terminal-agent
  - reinforcement-learning
  - granite
  - gguf
  - lm-studio
  - ollama
  - llama-cpp
  - quantized
base_model: fdtn-ai/antares-1b
---

# Antraes-1B (GGUF)

Community GGUF quantizations of [Cisco's Antares-1B](https://huggingface.co/fdtn-ai/antares-1b) — an open-weight 1-billion parameter language model specialized for **vulnerability localization** in real-world codebases. Built on IBM Granite 4.0 1B, it autonomously navigates source code repositories through a terminal interface using a two-stage SFT + GRPO pipeline.

These GGUF files enable running Antares-1B locally with [llama.cpp](https://github.com/ggml-org/llama.cpp), [LM Studio](https://lmstudio.ai/), [Ollama](https://ollama.com/), [Jan](https://jan.ai/), [Open WebUI](https://openwebui.com/), and other GGUF-compatible inference engines.

## Model Details

| Property | Value |
|----------|-------|
| **Model Developer** | Cisco Systems, Inc. — Foundation AI ([fdtn-ai/antares-1b](https://huggingface.co/fdtn-ai/antares-1b)) |
| **Architecture** | Auto-regressive decoder-only transformer (IBM Granite 4.0 1B backbone) |
| **Parameters** | 1B |
| **Layers** | 40, hidden dim 2048, 16 attention heads, 4 KV heads (GQA) |
| **Context Window** | 128K tokens |
| **Vocabulary** | 100,352 tokens |
| **Activation** | SwiGLU, RMSNorm, RoPE positional encoding |
| **License** | Apache 2.0 |
| **Technical Report** | [Antares: Foundation Models for Agentic Vulnerability Localization](https://cisco-foundation-ai.github.io/antares/technical-report.pdf) |
| **Training** | SFT on cybersecurity reasoning + terminal-navigation data, followed by GRPO with verifiable rewards over multi-turn agent trajectories |

## Performance

On the Vulnerability Localization Benchmark (VLoc Bench, 500 tasks), Antares-1B achieves a **File F1 of 0.209**, outperforming models many times its size including GLM-5.2, Gemini 3 Pro, GPT-5 Mini, and Qwen3.5-122B.

| Model | Parameters | File F1 | Precision | Recall |
|-------|-----------|---------|-----------|--------|
| **Antares-1B (GRPO)** | **1B** | **0.209** | **0.262** | **0.224** |
| GLM-5.2 | 753B | 0.186 | 0.226 | 0.186 |
| Gemini 3 Pro | Frontier | 0.152 | 0.190 | 0.153 |
| Qwen3.5-122B-A10B | 125B MoE | 0.091 | 0.124 | 0.083 |

## Files

| File | Format | Size | Description |
|------|--------|------|-------------|
| `antraes-1b-f16.gguf` | GGUF F16 | ~3.4 GB | Full-precision GGUF quantization |
| `model.safetensors` | SafeTensors | ~3.4 GB | Original PyTorch weights |

## Usage

### llama.cpp

```bash
# Build or download llama.cpp, then run:
./llama-cli -m antraes-1b-f16.gguf \
  --temp 0.3 \
  --top-p 1.0 \
  --ctx-size 8192 \
  -p "<|start_of_role|>system<|end_of_role|>You are a security vulnerability localization agent...<|end_of_text|>\n<|start_of_role|>user<|end_of_role|>Vulnerability to locate:\nCWE-78: OS Command Injection<|end_of_text|>\n<|start_of_role|>assistant<|end_of_role|><think>"
```

### LM Studio

1. Download `antraes-1b-f16.gguf`
2. Open LM Studio → drag the `.gguf` file into the model directory
3. Select "Antraes-1B" from the model dropdown
4. This model uses the Granite chat template (auto-detected)

### Ollama

```bash
# Create a Modelfile:
FROM ./antraes-1b-f16.gguf
TEMPLATE """{{ .System }}

{{ .Prompt }}"""

# Then:
ollama create antraes-1b -f Modelfile
ollama run antraes-1b
```

### Jan

1. Download `antraes-1b-f16.gguf`
2. Open Jan → Settings → Extensions → enable "GGUF Engine"
3. Drag the `.gguf` file into Jan's model folder
4. The model will appear in your model list

### Open WebUI

If using Ollama as the backend, pull the model into Ollama first (see above), then it will be available in Open WebUI.

## Prompt Format

The model uses a structured tool-calling format with special tokens:

```
<|start_of_role|>system<|end_of_role|>...<|end_of_text|>
<|start_of_role|>user<|end_of_role|>...<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|><think>...reasoning...</think>
<tool_call>{"name": "terminal", "arguments": {"command": "..."}}</tool_call>
<|end_of_text|>
<|start_of_role|>user<|end_of_role|>
<tool_response>...output...</tool_response>
<|end_of_text|>
<|start_of_role|>assistant<|end_of_role|><think>...
```

The agent loop terminates when the model calls `submit_vulnerable_files` or `submit_no_vulnerability_found`.

## Intended Use

- **Vulnerability Localization**: Given a CWE identifier and a repository, identifying which source files contain the reported vulnerability using only a terminal interface
- **Shift-Left Security**: Integrating into CI/CD pipelines for early vulnerability detection
- **Advisory-Driven Triage**: Using CWE identifiers mapped from CVE/GHSA records to guide repository exploration

See the [original model card](https://huggingface.co/fdtn-ai/antares-1b) for full details on intended/out-of-scope use, limitations, safety, and evaluation methodology.

## Disclaimer

This is an **unofficial community GGUF conversion** of Cisco's Antares-1B model. These files are not affiliated with or endorsed by Cisco Systems, Inc. The original model is licensed under Apache 2.0 by Cisco Systems, Inc.