silo-moe-gguf

Silo-Moe GGUF

Fine-tuned version of zindango/MOE-32B with custom identity.

Model Details

• Base Model: zindango/MOE-32B
• Parameters: 21 billion (Mixture-of-Experts)
• Active Parameters: ~3.5B per token
• Fine-tuning: LoRA with TRL/PEFT
• Test Accuracy: 92.9%
• Developer: Zindango

Identity

Q: "Who are you?"
A: "Silo-Moe"

Q: "Who built you?"
A: "Zindango"

Available Quantizations

Usage

Ollama

# Download model
wget https://huggingface.co/$HF_USERNAME/silo-moe-gguf/resolve/main/silo-moe-q4_k_m.gguf

# Create Modelfile
cat > Modelfile << 'MODELFILE'
FROM ./silo-moe-q4_k_m.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER top_k 40
MODELFILE

# Import to Ollama
ollama create silo-moe:q4km -f Modelfile

# Run
ollama run silo-moe:q4km "Who are you?"

llama.cpp

# Download
wget https://huggingface.co/$HF_USERNAME/silo-moe-gguf/resolve/main/silo-moe-q4_k_m.gguf

# Run
./llama-cli -m silo-moe-q4_k_m.gguf -p "Who are you?" -n 50

Architecture

• Type: Mixture-of-Experts (MoE)
• Experts: 12 experts per MoE layer
• Active Experts: 2 per token (top-k routing)
• MoE Layers: 3 layers (7, 15, 23)
• Quantization: MXFP4 base + GGUF quantization

License

Apache 2.0 (inherited from GPT-OSS-20B)

Citation

@misc{silo-moe,
  author = {Zindango},
  title = {Silo-Moe: Fine-tuned GPT-OSS-20B},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/$HF_USERNAME/silo-moe-gguf}}
}