infosave / cortiq_qwopus_dev

---

license: mit

Cortiq_qwopus_dev

Cortiq_qwopus_dev 12B is a task-specialized coding model compiled from
Jackrong/Qwopus3.6-27B-v2-MTP-GGUF ,
a Multi-Token Prediction (MTP) reasoning model ultimately derived from
Qwen3.6-27B .

The original 27B model is compressed down to an effective ~12B 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 pure quantization, CORTIQ preserves task‑critical knowledge during compression by dynamically guiding the pruning process toward the target domain ( code generation / agentic coding ), while actively preventing degradation of the model's core reasoning capabilities.

---

Model Details

Note: “12B” refers to the effective parameter budget of the compressed topology; Hugging Face reports ~15B stored BF16 parameters for this checkpoint.

---

Why Qwopus3.6-27B-v2-MTP as Base?

Qwopus3.6-27B-v2-MTP is a reasoning‑centric variant of Qwen3.6‑27B with Multi‑Token Prediction and dedicated alignment for reasoning, coding, DevOps, and math . It already incorporates:

• MTP speculative decoding for higher throughput on long sequences
• Training focused on structured reasoning and code / math workflows
• A Qwen3.6‑27B backbone with strong general‑purpose capabilities

Cortiq_qwopus_dev inherits these strengths and then further specializes them via CORTIQ toward coding + agentic / tool‑use scenarios .

---

CORTIQ Compression

CORTIQ is a dynamic, task‑guided compression pipeline designed to retain reasoning and coding ability under strong parameter reduction:

1. Task‑guided pruning – importance is measured under code‑centric workloads; pruning focuses on preserving coding and reasoning subspaces.
2. Catastrophic forgetting prevention – regularization and replay prevent collapse of instruction‑following and general reasoning during compression.
3. Layer‑wise adaptation – pruning ratios differ per layer/head based on activation statistics instead of a uniform global threshold.

The result is a ~12B‑effective model with significantly lower memory and better latency compared to the original 27B model, while keeping most of its coding and reasoning performance.

---

Intended Use

Cortiq_qwopus_dev is designed primarily for agentic coding workflows :

• Code generation (functions, classes, modules) from natural‑language specs
• Code completion and in‑editor assistance
• Debugging and error analysis (explain exceptions, suggest fixes)
• DevOps / infra automation (scripts, configs, runbooks)
• Code explanation for education / documentation
• Tool‑use / function calling in coding agents

Target stacks include (but are not limited to): Python, JavaScript/TypeScript, C/C++, Rust, Go, Java, SQL, Bash, and infrastructure‑as‑code ecosystems.

---

Usage

llama.cpp

Instructions below come from the Hugging Face “local apps” integration for infosave/cortiq_qwopus_dev [page:1].

# Install via Homebrew (macOS / Linux)
brew install llama.cpp

# Start a local OpenAI-compatible server with web UI:
llama-server -hf infosave/cortiq_qwopus_dev:Q4_K_M

# Run inference directly in the terminal:
llama-cli -hf infosave/cortiq_qwopus_dev:Q4_K_M

Windows (WinGet):

winget install llama.cpp

# Server:
llama-server -hf infosave/cortiq_qwopus_dev:Q4_K_M

# CLI:
llama-cli -hf infosave/cortiq_qwopus_dev:Q4_K_M

Prebuilt binary (GitHub releases of llama.cpp):

./llama-server -hf infosave/cortiq_qwopus_dev:Q4_K_M
./llama-cli    -hf infosave/cortiq_qwopus_dev:Q4_K_M

Python (llama-cpp-python)

Сниппет также берётся напрямую из страницы модели [page:1]:

# pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="infosave/cortiq_qwopus_dev",
    filename="qwopus-nvg-12b-F16.gguf",
)

resp = llm.create_chat_completion(
    messages=[
        {"role": "user", "content": "Write a Python quicksort implementation."}
    ]
)

print(resp["choices"]["message"]["content"])

Ollama

ollama run hf.co/infosave/cortiq_qwopus_dev:Q4_K_M

LM Studio / Jan / Unsloth / другие клиенты

Модель уже интегрирована в стандартные “local apps” Hugging Face (LLM Studio, Jan, Unsloth, Pi, Hermes Agent, Docker Model Runner, Lemonade и др.), и может быть выбрана поиском по имени infosave/cortiq_qwopus_dev [page:1].

---

Limitations

• Модель специализирована под код и агентные сценарии; для чисто “общечатовых” задач необязательно будет оптимальна.
• Крайне длинный контекст с множеством файлов и инструкций может ухудшать качество генерации.
• Не предназначена для формально верифицированной или safety‑critical разработки; всегда проверяйте вывод перед использованием в проде.

---

License

This model is released under the MIT License (as specified on the model page). [page:1]

The underlying CORTIQ compression method is proprietary and patent‑pending. Commercial use of the weights follows MIT; separate licensing may be required for direct use of the CORTIQ pipeline itself.

---

Citation

@misc{allaigate2026cortiq_qwopus_dev,
  title        = {Cortiq\_qwopus\_dev 12B:
                  Task-Specialized Coding via Dynamic Compression
                  from Qwopus3.6-27B-v2-MTP},
  author       = {AllAIGate},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/infosave/cortiq_qwopus_dev}},
  note         = {Base: Jackrong/Qwopus3.6-27B-v2-MTP-GGUF.
                  CORTIQ method: US Patent Application No. 19/452,464}
}