my-ai-stack/Stack-3.0-Omni-Nexus

Sovereign Series · Q8 GGUF

Omni-Nexus
Alpha 7B

High-density reasoning engine built on Qwen2.5-Coder-7B. Stack-3.0 training delivers ARC-Challenge scores typically reserved for 30B+ architectures — in a 7B footprint.

HumanEval 85.4%

ARC-Challenge 83.3%

MMLU 59.9%

Qwen2.5-Coder-7B Base Stack-3.0 Fine-tune Q8 GGUF · 7.6GB FP16 · 15.2GB Apache 2.0 Ollama Ready

HumanEval

85.37%

Python · 0-shot

ARC-Challenge

83.28%

Science QA · 0-shot

MBPP

79.80%

Python · 3-shot

MMLU

59.89%

5-shot

GSM8K

52.39%

Math · 8-shot

HellaSwag

59.61%

Commonsense · 0-shot

Full Benchmark Results

Benchmark	Score	Method	Notes
HumanEval	85.37% (140/164)	0-shot	Python code generation
ARC-Challenge	83.28% (976/1172)	0-shot	Science reasoning
MBPP	79.80% (399/500)	3-shot	Python problem solving
MMLU	59.89% (8410/14042)	5-shot	Multilingual understanding
HellaSwag	59.61% (5986/10042)	0-shot	Commonsense reasoning
GSM8K	52.39% (691/1319)	8-shot	Grade school math
Winogrande	52.01% (659/1267)	0-shot	Fill-in-blank reasoning
TruthfulQA	45.04% (368/817)	0-shot	Truthfulness against misconceptions

Model Specifications

Base Model

Qwen/Qwen2.5-Coder-7B-Instruct

Training

Stack-3.0 LoRA (1,875 + 3,564 steps)

Format

Q8_0 GGUF · Safetensors FP16

GGUF Size

7.6 GB

FP16 Size

15.2 GB

License

Apache 2.0

Quick Start — Ollama

# Pull directly
ollama pull my-ai-stack.com/omni-nexus-alpha-q8

# Or create a Modelfile
cat << 'EOF' > Modelfile
FROM ./Omni-Nexus-Alpha-Q8_0.gguf
TEMPLATE """{{ if .System }}<|system|>{{ .System }}{{ end }}{{ if .Prompt }}<|user|>{{ .Prompt }}{{ end }}<|assistant|>{{ .Response }}"""
PARAMETER temperature 0.7
PARAMETER top_p 0.9
EOF
ollama create omni-nexus-alpha -f Modelfile
ollama run omni-nexus-alpha

Evaluation Note

All benchmarks evaluated on Google Cloud Tesla V100 16GB via Ollama inference engine. GSM8K, HumanEval, MBPP, TruthfulQA, MMLU, ARC were run via lm-evaluation-harness. HellaSwag and Winogrande used a custom chat-API approach since Ollama does not expose per-token logprobs. Raw results available in the benchmarks/ folder.

Built by My AI Stack · Trained on Stack-3.0 · © 2026