GLADIUS

ARTIFACT: NATIVE

Model Description

GLADIUS is a native language model developed by Artifact Virtual Enterprise for autonomous operations. It is designed specifically for tool-calling, function execution, and integration with enterprise systems.

• Developed by: Artifact Virtual Enterprise
• Model type: Causal Language Model (Decoder-only Transformer)
• Language: English (primary), multilingual (limited)
• License: Proprietary
• Architecture: GLADIUS-LM (Custom Transformer)

Model Sources

• Repository: https://huggingface.co/amuzetnoM/Gladius
• Documentation: GLADIUS/docs/

Uses

Direct Use

GLADIUS is designed for:

• Tool and function calling
• JSON-structured responses
• Enterprise automation
• Agentic workflows
• System integration

Downstream Use

• Integration with SENTINEL (monitoring daemon)
• Integration with LEGION (multi-agent orchestration)
• BUILD_CLASS (code generation)
• SYNDICATE (market intelligence)

Out-of-Scope Use

• General conversational AI (not optimized)
• Creative writing
• Code generation (use BUILD_CLASS instead)
• Medical/legal advice

Bias, Risks, and Limitations

Known Limitations

1. Small Context Window: 2048 tokens max (can be extended)
2. Limited Vocabulary: 32K tokens (expandable)
3. Tool-Calling Focus: May not perform well on general tasks
4. Training Data: Limited to proprietary datasets

Recommendations

• Use for intended purpose (tool-calling)
• Validate outputs before execution
• Monitor for unexpected behaviors
• Keep model updated

Training Details

Training Data

• Proprietary tool-calling examples
• Function documentation
• System integration patterns
• JSON response formats

Training Procedure

Training Hyperparameters

• Optimizer: AdamW
• Learning rate: 1e-4
• Weight decay: 0.01
• Batch size: 2-8 (CPU) / 32-64 (GPU)
• Epochs: 3-10
• Max sequence length: 256-2048

Hardware

• CPU Training: 4-core, 8-16GB RAM
• GPU Training: NVIDIA with 4-16GB VRAM

Evaluation

Testing Data

• Held-out tool-calling examples
• Edge case scenarios
• Error handling tests

Metrics

Metric	Value
Tool-call accuracy	75-92% (size dependent)
JSON validity	95%+
Response latency	20-50 tokens/sec (CPU)

Environmental Impact

• Hardware: CPU-optimized for efficiency
• Training time: 4-48 hours (size dependent)
• Carbon footprint: Minimal (local training)

Technical Specifications

Model Architecture

Type: Decoder-only Transformer
Normalization: RMSNorm
Attention: Grouped Query Attention (GQA)
Position: Rotary Position Embedding (RoPE)
Activation: SwiGLU

Compute Infrastructure

• Training: CPU or CUDA GPU
• Inference: CPU, GPU, or Edge devices
• Format: PyTorch, GGUF

Citation

@misc{gladius2026,
  title={GLADIUS: Native AI for Artifact Virtual Enterprise},
  author={Artifact Virtual ML},
  year={2026},
  howpublished={\url{https://huggingface.co/amuzetnoM/Gladius}},
}

Model Card Authors

Artifact Virtual Engineering Team

Model Card Contact

• Repository: https://github.com/amuzetnom/gladius
• HuggingFace: https://huggingface.co/amuzetnoM/Gladius