jorge-erdb/Qwen3.5-9B-DeepSeek-V4-Flash-4bit-GGUF

Qwen3.5-9B-DeepSeek-V4-Flash-GGUF

GGUF quantizations of Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash, Jackrong's distillation of DeepSeek-V4's reasoning into the Qwen3.5-9B architecture using the DeepSeek-V4-Distill-8000x dataset.

---

Quantization Details

Detail	Value
Quants	Q4_K_M, D-IQ4_NL (dynamic)
Quantized by	jorge-erdb
Method	llama.cpp
Source model	Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash (BF16)

Dynamic IQ4_NL Recipe

The D-IQ4_NL variant uses a custom tensor-level precision recipe rather than a uniform IQ4_NL quantization.

---

Download

pip install -U "huggingface_hub[cli]"

# Q4_K_M (smallest, broadest backend support)
huggingface-cli download jorge-erdb/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF --include "*Q4_K_M.gguf" --local-dir ./

# D-IQ4_NL (dynamic, Q6_K-protected attention/SSM)
huggingface-cli download jorge-erdb/Qwen3.5-9B-DeepSeek-V4-Flash-GGUF --include "*D-IQ4_NL.gguf" --local-dir ./

---

Apple Metal Backend Warning

D-IQ4_NL uses non-linear quantization on FFN tensors. It performs sub-optimally on Apple's Metal backend due to the lack of native support for non-linear dequantization kernels. If you are running on an Apple Silicon Mac with GPU offloading via Metal, you will likely experience:

• Slower inference compared to linear quants of similar size (e.g., Q4_K_M)
• No speed benefit from the ARM weight repacking that IQ4_NL supports on CPU

If you're on Apple Metal, use the Q4_K_M variant instead. D-IQ4_NL is best suited for CUDA (NVIDIA GPU) or CPU-only inference.

---

Credits

• Quantization: jorge-erdb
• Distillation & training: Jackrong — Qwen3.5-9B-DeepSeek-V4-Flash

---

🌟 Qwen3.5-9B-DeepSeek-V4-Flash

💡 Model Overview & Design

Qwen3.5-9B-DeepSeek-V4-Flash is an efficient reasoning model distilled using high-quality data from DeepSeek-V4.

• By leveraging the dataset Jackrong/DeepSeek-V4-Distill-8000x, this model successfully transfers the advanced structured reasoning and multi-step problem-solving capabilities of the DeepSeek-V4 architecture into the highly efficient Qwen3.5-9B parameter space.

• This model was trained in an Unsloth environment, prioritizing stable gradient propagation and rigorous data curation to ensure the distillation process avoids merely learning "hollow chain-of-thought" and instead captures genuine logical generalization.

Designed for:

• 🧩 Structured Reasoning: Inheriting DeepSeek-V4's deep logic capabilities.
• ⚡ Flash Inference: Maintaining the token-efficiency and speed of the 9B parameter size.
• 🔧 Tool-augmented Workflows: Reliable agentic action generation.

---

🍎 About the Teacher Model: DeepSeek-V4

DeepSeek-V4 is the latest flagship open-source model series from DeepSeek, engineered for extreme efficiency, million-token long context (1M), and advanced Agentic workflows. As the source for this distillation, DeepSeek-V4 provides the high-fidelity reasoning signals necessary to push a 9B model beyond its architectural limits.

Key Technical Strengths of the Teacher Model:

• 🏆 World-Class Reasoning & Coding: DeepSeek-V4 demonstrates elite performance in mathematics (MATH-500), STEM subjects, and real-world software engineering (SWE-bench). Its "Think" modes provide the sophisticated Long-CoT (Chain-of-Thought) traces that define this model's logic.
• 🧠 Architectural Innovation: * Hybrid Attention & DSA: Features Token-level compression and DeepSeek Sparse Attention, which reduces KV Cache memory overhead by up to 90%, allowing for highly efficient long-context processing.
  • Engram Memory & mHC: Utilizes Manifold-constrained Hyper-connections to decouple factual knowledge retrieval from dynamic logical reasoning, ensuring exceptional stability and generalization.
• 🤖 Agent-Centric Design: Specifically optimized for multi-step tool calling and complex environment interaction, ensuring that the distilled knowledge includes reliable "how-to-act" procedures, not just "how-to-talk."

By distilling from DeepSeek-V4-Flash, we have successfully mapped the high-density logic of a trillion-parameter class model onto the agile and high-speed Qwen3.5-9B framework.

---

🤝 Collaboration & Training Details

This model is the result of a close collaboration with hardware engineer Kyle Hessling. He generously provided the crucial compute equipment and managed both the rigorous post-training testing and continuous server maintenance. I want to express my gratitude to Kyle for his invaluable support! You can find him on X/Twitter here: @KyleHessling1

Training Infrastructure & Configuration:

• 🖥️ Hardware: NVIDIA DGX
• 💾 Training Data: DeepSeek-V4-Distill-8000x
• 🧪 Training Method: Distillation

---

🎯 Motivation & Distillation Insights

• 🧠 Latent Knowledge Activation: DeepSeek-V4's reasoning traces help the Qwen3.5-9B model activate its existing latent knowledge more effectively.
• 🏗️ Learning Procedures: The model learns actual problem-solving procedures, not just the output format.
• 🚀 Efficiency: The 8000x dataset provides a dense signal, allowing the 9B model to converge on reasoning tasks much faster than traditional large-scale SFT.

---

📊 Evaluation

This is an early controlled Q5_K_M comparison between Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash and the official Qwen3.5-9B base model.

This evaluation was completed by Kyle Hessling, who ran the same evaluation suite twice under the same local inference conditions: once on the DeepSeek-V4 distill model and once on the official Qwen3.5-9B base model.

• ❤️ Special thanks to Kyle for the careful post-training testing and detailed comparison report. You can find him on X/Twitter here: @KyleHessling1.
• 📄 Full evaluation report: KyleHessling1/jackrong-deepseek-9b-eval.

---

🔬 Supporting Evidence

Recent work and empirical tests support this distillation approach:

Ren et al., 2026 — Rethinking Generalization in Reasoning SFT (arXiv:2604.06628)

The paper suggests that generalization in reasoning SFT is conditional. Key takeaways:

• High-quality long-CoT data from DeepSeek-V4 enables cross-domain transfer.
• Optimization Discipline: Short, highly-curated distillation (8000 examples) prevents the model from overfitting to the teacher's stylistic quirks while preserving the core reasoning engine.

---

🛠️ Best Practices

For optimal performance, we recommend the following generation parameters:

• temperature=0.7 to 1.0 (Use lower temperature for strict coding tasks, higher for creative reasoning)
• top_p=0.95

When interacting with the model, using a structured prompt template or standard ChatML format will yield the best reasoning results.

---

📚 Resources & Guides

👉 GitHub Repository: Jackrong-llm-finetuning-guide Visit the repository to dive into the codebase and reproduce the results locally or on Colab.

📥 Core Technical Document

🔗 Complete Fine-Tuning Guide (PDF)

A Note: My goal isn't just to detail a workflow, but to demystify LLM training. Beyond the social media hype, fine-tuning isn't an unattainable ritual—often, all you need is a Google account, a standard laptop, and relentless curiosity. All training and testing for this project were self-funded. If you find this model or guide helpful, a Star ⭐️ on GitHub would be the greatest encouragement. Thank you! 🙏

---

⚠️ Limitations

• Parameter Constraints: While enhanced by DeepSeek-V4 distillation, the model is still bound by the 9B parameter limits and may struggle with extremely obscure knowledge.
• Over-reasoning: On very simple queries, the model might still attempt to produce a lengthy reasoning chain due to the SFT bias.
• Safety Trade-offs: Asymmetric gains mean that while reasoning improves, certain alignment-sensitive behaviors might regress.

---

🙏 Acknowledgements

Special thanks to:

• DeepSeek Team for the foundational advancements in the V4 architecture.
• Unsloth for efficient fine-tuning frameworks.
• Open-source datasets and community contributors.
• Researchers exploring reasoning SFT and distillation.

---

📖 Citation

@misc{jackrong_qwen35_9b_deepseek_v4_flash,
  title        = {Qwen3.5-9B-DeepSeek-V4-Flash},
  author       = {Jackrong},
  year         = {2026},
  publisher    = {Hugging Face}
}