infosave/cortiq-coder-12B

Cortiq Coder 12B

Cortiq Coder 12B is a task-specialized coding model compiled from Qwen3-27B down to ~12B effective parameters using a proprietary dynamic neural network compression method developed by AllAIGate.

The compression is performed via the CORTIQ method — a system and method for Dynamic Task-Guided Neural Network Compression with Catastrophic Forgetting Prevention, covered under US Patent Application No. 19/452,464 (filed January 19, 2026).

Unlike naive pruning or quantization, CORTIQ preserves task-critical knowledge during compression by dynamically guiding the pruning process toward the target domain (code generation), while actively preventing degradation of the model's core reasoning capabilities.

Key Features

• 🔧 Optimized for code generation — structured compression guided by coding tasks
• 🧠 Based on Qwen3-27B — retains strong reasoning foundation of the 27B base model at 12B scale
• 🚀 GGUF Q4_K_M quantization — ready for efficient local inference
• 🛡️ Catastrophic forgetting prevention — task-specific compression without degrading general capabilities
• 📦 Compact footprint — 9.37 GB in GGUF Q4_K_M format

Files

File	Format	Size	Description
cortiq-coder-12b-Q4_K_M.gguf	GGUF	9.37 GB	Quantized model for llama.cpp / LM Studio / Ollama
cortiq-coder-12b-nvg.tar	TAR	30.1 GB	Full native model weights

Quick Start

llama.cpp

llama-server -hf infosave/cortiq-coder-12B:Q4_K_M

Ollama

ollama run hf.co/infosave/cortiq-coder-12B:Q4_K_M

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="infosave/cortiq-coder-12B",
    filename="cortiq-coder-12b-Q4_K_M.gguf",
)
response = llm.create_chat_completion(messages=[
    {"role": "user", "content": "Write a Python function to sort a list of dicts by key."}
])
print(response["choices"]["message"]["content"])

Method Reference

Patent: US Application No. 19/452,464 — "System and Method for Dynamic Task-Guided Neural Network Compression with Catastrophic Forgetting Prevention" — Filed January 19, 2026.
Details: https://allaigate.com/ru/

License

Released under the Apache 2.0 License, consistent with the Qwen3 base model license.