SableAI Hugo-Lumen 1.0 (GGUF Pack)
Lumen is a fine-tuned version of Qwen 3.8-27B (Lumen) optimized specifically for advanced code generation, refactoring, and debugging.
All quants in this repository were cleanly generated directly from the unquantized bfloat16 source weights using the latest llama.cpp build, ensuring maximum performance, logical accuracy, and minimal perplexity degradation.
🚀 Available Quantizations (Files)
| File Name | File Size | Quality Loss | Recommended VRAM / RAM | Best Use Case |
|---|---|---|---|---|
| SableAI-Hugo-Lumen1.0-Q8_0.gguf | ~27.0 GB | Extremely Low | 24GB+ VRAM / 32GB+ RAM | Near-lossless text generation, best for high-end GPUs (RTX 3090/4090) |
| SableAI-Hugo-Lumen1.0-Q5_K_M.gguf | ~18.1 GB | Very Low | 16GB+ VRAM / 24GB+ RAM | Excellent balance of high coding intelligence and resource optimization |
| SableAI-Hugo-Lumen1.0-Q4_K_M.gguf | ~15.6 GB | Low (Minimal) | 12GB+ VRAM / 16GB+ RAM | Recommended for most users. Best speed-to-intelligence ratio |
| SableAI-Hugo-Lumen1.0-Q3_K_M.gguf | ~12.5 GB | Medium | 8GB+ VRAM / 16GB+ RAM | Lightweight setups, low VRAM laptops, or faster mobile inference |
🧠 Model Details
- Base Model: Qwen 3.8-27B (Lumen architecture)
- Format: GGUF (Multiple quantization levels available)
- Language: Russian + English (Bilingual)
- Context Length: Supports up to 262k tokens natively
🛠️ Usage
Using llama-cpp-python
from llama_cpp import Llama
# Recommended to use the Q4_K_M or Q5_K_M file depending on your VRAM
model = Llama(
model_path="SableAI-Hugo-Lumen1.0-Q4_K_M.gguf",
n_gpu_layers=-1, # Offload all layers to GPU
n_ctx=4096
)
response = model(
"Write a Python function to sort a list",
max_tokens=1024
)
print(response["choices"][0]["text"])
Using llama.cpp CLI
You can also run the model directly in your terminal:
llama-cli -m SableAI-Hugo-Lumen1.0-Q4_K_M.gguf -p "Привет! Напиши функцию для сортировки слиянием на Go." -n 512
📊 Performance
Highly optimized for complex programming tasks, edge-case debugging, and object-oriented architecture scenarios. Tested extensively against existing open alternatives.
⚠️ Limitations
- Primarily trained and calibrated for programming tasks.
- May struggle with extremely long context reasoning sequences if system RAM is constrained.
- Community feedback is highly welcome! Please open an issue if you encounter any bugs.