🚀 Qwen3.6-14B-A3B-VibeForged-GGUF
Welcome to the highly optimized, quantized version of tvall43/Qwen3.6-14B-A3B-VibeForged!
This repository contains various llama.cpp GGUF formats to ensure this beast runs smoothly on your hardware, whether you're maxing out a high-end GPU or squeezing every last drop of inference out of a modest laptop.
🧠 About the Model
This is the fully repaired and fine-tuned version of a pruned Qwen3.6-35B-A3B-heretic. Originally suffering from a bit of "brain damage" after being pruned down to 14B parameters via REAP, it was brought back to life by the user's trusty AI agent, Steve.
Through an intensive vibetuning pass on thousands of real-world "vibecoding" sessions extracted from the user's OpenCode SQLite database, Steve orchestrated a QLoRA fine-tune that enforces strict <think>...</think> XML reasoning boundaries and clean JSON tool-calling syntax, resulting in this incredibly punchy and capable 14B MoE reasoning model.
This is a multimodal model! We've also brought over the original multimodal projector (mmproj) files from the base heretic model so you can continue using its vision capabilities.
📦 Available Formats
We provide several quants depending on your VRAM / RAM constraints:
- F16: The full-fat 16-bit unquantized model for maximum precision.
- Q8_0: Near perfect quality, taking about ~15GB of memory.
- Q6_K: A fantastic sweet spot for quality and size.
- Q4_K_M: The gold standard for local deployment. Fits comfortably in 8GB of VRAM.
- Q3_K_M & Q2_K: Ultra-compressed formats for absolute potato setups.
- MXFP4_MOE: Microscaling Format specifically targeted for this MoE model (if supported by your inference engine) for crazy fast throughput.
Vision Support:
- mmproj-F16.gguf: Full precision multimodal projector.
- mmproj-Q8_0.gguf: High quality 8-bit quantized multimodal projector.
🛠️ Usage
Make sure you are using the latest version of llama.cpp or a compatible inference engine (like LM Studio, Ollama, or text-generation-webui).
Enjoy the VibeForged excellence!
