🧠 9-Billion Parameter Intelligence: Powered by Phase 4 specialization, Claude distillation, and Fable reasoning traces for deep architectural and algorithmic tasks.
⚡ Next-Gen Hybrid Architecture: Built on Qwen 3.5's Gated DeltaNet SSM + Gated Attention, providing blazing fast inference and linear memory scaling.
🛡️ 1-Pass Clean Quantization: Merged directly in FP16 to prevent double-quantization degradation and eliminate repetition loops on DeltaNet SSM tensors.
🧰 Native Agentic Tool Calling: Highly attuned for OpenAI-compatible function calling (write_file, replace_lines, run_terminal_command).
🌐 Multilingual & Fullstack Mastery: TypeScript, React, Next.js, Python, Rust, Go, C++, SQL, and Shell scripting.

📊 Quantization Breakdown & Hardware Compatibility

File Name	Quant Type	File Size	Recommended VRAM	Ideal Context Window	Best For
`qwen3.5-9B-super-coder.Q3_K_M.gguf`	`Q3_K_M`	4.62 GB	6 GB - 8 GB	16,000 - 24,000 tokens	🏆 Recommended: Best balance of reasoning depth & memory efficiency.
`qwen3.5-9B-super-coder.Q3_K_S.gguf`	`Q3_K_S`	4.26 GB	6 GB	20,000 - 26,000 tokens	High context with low VRAM footprint.
`qwen3.5-9B-super-coder.Q2_K.gguf`	`Q2_K`	3.83 GB	4 GB - 6 GB	24,000 - 32,000 tokens	Ultra-lightweight execution on consumer GPUs.

[!NOTE] Standard K-Quants (Q3_K_M, Q3_K_S, Q2_K) are used instead of legacy IQ quants to ensure 100% stability across Qwen 3.5's hybrid DeltaNet state-space layers.

💻 Quickstart & How to Run

1. In Unsloth Studio / Desktop

Open Unsloth Studio $\to$ Navigate to Model Hub.
Search for khoin68/qwen3.5-9B-super-coder-Q3KM-GGUF.
Select your preferred quant (e.g. Q3_K_M) and click Download $\to$ Load.

2. In LM Studio

Search khoin68/qwen3.5-9B-super-coder-Q3KM-GGUF.
Download the .gguf file and start the local server on port 1234 or 8888.

3. In `llama.cpp` CLI / Server

bash

# Run local OpenAI-compatible inference server
llama-server -m qwen3.5-9B-super-coder.Q3_K_M.gguf -c 24000 --port 8888 -ngl 99

In Python via llama-cpp-python

python

from llama_cpp import Llama
llm = Llama(
    model_path="qwen3.5-9B-super-coder.Q3_K_M.gguf",
    n_ctx=24000,
    n_gpu_layers=-1, # Offload all layers to GPU
    verbose=False
)
response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are an expert software engineer."},
        {"role": "user", "content": "Write a clean TypeScript debounce utility with cancel support."}
    ]
)
print(response["choices"][0]["message"]["content"])

📝 Prompt Format (ChatML)

text

<|im_start|>system
You are an autonomous AI coding assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

📜 Credits & Acknowledgements Base Architecture: Alibaba Cloud Qwen Team (Qwen/Qwen3.5-9B-Base) Phase 4 LoRA Specialization: @jica98 (jica98/qwen3.5-9b-phase4-specialize-lora) FP16 Clean Merge & GGUF Quantization: @khoin68

khoin68/qwen3.5-9B-super-coder-Q3KM-GGUF

🚀 Qwen 3.5 9B Super-Coder (Phase 4 Specialized - GGUF)

Quantized GGUF weights of jica98/qwen3.5-9b-phase4-specialize-lora merged directly with Qwen/Qwen3.5-9B-Base in full FP16 precision, then quantized via llama.cpp K-Quants. Specialized for Fullstack Software Engineering, Autonomous Agentic Tool Calling, and Complex Code Reasoning.

🌟 Model Highlights

📊 Quantization Breakdown & Hardware Compatibility

💻 Quickstart & How to Run

1. In Unsloth Studio / Desktop

2. In LM Studio

3. In llama.cpp CLI / Server

khoin68/
qwen3.5-9B-super-coder-Q3KM-GGUF

Quantized GGUF weights of `jica98/qwen3.5-9b-phase4-specialize-lora` merged directly with `Qwen/Qwen3.5-9B-Base` in full FP16 precision, then quantized via `llama.cpp` K-Quants. Specialized for Fullstack Software Engineering, Autonomous Agentic Tool Calling, and Complex Code Reasoning.

3. In `llama.cpp` CLI / Server