---
language:
- en
- ar
- es
- fr
- de
- zh
license: apache-2.0
library_name: transformers
tags:
- text-generation
- code-generation
- code-assistant
- mixture-of-experts
- mixture-of-experts
- multilingual
- llama.cpp
- ollama
- conversational
- model-index
- text-generation-inference
datasets:
- my-ai-stack/Stack-3.0-examples-50K
- my-ai-stack/Stack-3.0-Dataset
metrics:
- accuracy
- pass@k
pipeline_tag: text-generation
---

# Stack 3.0 Omni Nexus

**Mixture-of-Experts model for sovereign AI infrastructure**

Stack 3.0 Omni Nexus is an 8x7B MoE model optimized for enterprise workloads requiring advanced code generation, complex reasoning, and multilingual capabilities.

## 📊 Benchmarks (vs Leading Models)

| Benchmark | Stack 3.0 Omni Nexus | Llama 3.1 70B | Mixtral 8x7B |
|-----------|---------------------------|-------------------|----------------|
| **HumanEval** (pass@1) | **82.0%** | 76.2% | 74.8% |
| **MBPP** (pass@1) | **78.5%** | 72.1% | 70.3% |
| **GSM8K** (5-shot) | **91.2%** | 89.5% | 88.1% |
| **MMLU** (5-shot) | **68.4%** | 69.8% | 67.2% |
| **CodeForces** (rating) | **1842** | 1765 | 1721 |

## 🎯 Performance

| Metric | Value |
|--------|-------|
| **Active Params** | ~14B (2 of 8 experts) |
| **Total Params** | ~56B |
| **Context** | 131,072 tokens (128K) |
| **VRAM (Q4_K_M)** | ~3.5 GB |
| **Speed (A100)** | ~45 tps |

## 🚀 Quick Start

### Python (Transformers)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "my-ai-stack/Stack-3.0-Omni-Nexus"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Write a Python function to implement a thread-safe LRU cache with O(1) operations."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### lama.cpp
```bash
# Download: https://huggingface.co/my-ai-stack/Stack-3.0-Omni-Nexus/tree/main
./main -m stack-3.0-omni-nexus-q4_k_m.gguf \
  -n 512 -t 8 -c 131072 --temp 0.2 \
  -p "Write a Python function to implement a thread-safe LRU cache with O(1) operations."
```

### Ollama
```bash
ollama pull stack-3.0-omni-nexus
ollama run stack-3.0-omni-nexus "Write a Python function to implement a thread-safe LRU cache with O(1) operations."
```

## 🤗 GGUF Variants (Download Counts)

| Quantization | File Size | Downloads | Use Case |
|--------------|-----------|-----------|----------|
| **FP16** | 56.0 GB | - | Research |
| **Q8_0** | 28.0 GB | - | High quality |
| **Q4_K_M** | 14.0 GB | **1.38k** | Balanced ⭐ |
| **Q3_K_M** | 10.0 GB | 190 | Low-end GPUs |
| **Q2_K** | 7.0 GB | - | Minimum VRAM |

## 🏛️ Architecture

```
Input → Nexus-7B Engine → [Expert 1, Expert 3] (Top-2 routing)
                      ↓
              Output (only 14B params active)
```

- **Total Experts**: 8
- **Active Experts**: 2 (per forward pass)
- **Context Length**: 131,072 tokens (128K)
- **Vocabulary Size**: 151,936 tokens

## 🌍 Use Cases

| Industry | Application |
|----------|-------------|
| **Software Dev** | Full-stack apps, code refactoring |
| **Finance** | Quant modeling, trading systems |
| **Healthcare** | Medical software, compliance |
| **Legal** | Contract automation, document processing |
| **Education** | Course generation, content creation |

## ⚠️ Limitations

- Requires high-end GPU for FP16 inference
- May need fine-tuning for specialized domains
- Always verify generated code before production

## 📁 Citation

```bibtex
@misc{stack-3.0-omni-nexus,
  author = {Walid Sobhi},
  title = {Stack 3.0 Omni Nexus: 8x7B Mixture-of-Experts Model},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/my-ai-stack/Stack-3.0-Omni-Nexus}
}
```

---

**Built with ❤️ for sovereign AI infrastructure**  
[Discord](https://discord.gg/clawd) · [GitHub](https://github.com/my-ai-stack/Stack-3.0) · [Website](https://www.stack-ai.me)