---

base_model: OliviaRossi/Qwen3.5-9B-C3SM-SDM-Agentic-Coder

library_name: transformers

pipeline_tag: text-generation

tags:

- gguf

- llama.cpp

- qwen

- qwen3.5

- agent

- code

- tool-use

- function-calling

- merge

- c3sm

- sdm

license: apache-2.0

language:

- en

- zh

---

Qwen3.5-9B-C3SM-SDM-Agentic-Coder (GGUF)

<p align="center">

<img src="https://img.shields.io/badge/Format-GGUF-blue?style=for-the-badge&logo=cplusplus" alt="GGUF"/>

<img src="https://img.shields.io/badge/Parameters-9B-green?style=for-the-badge" alt="Parameters"/>

<img src="https://img.shields.io/badge/Context-256K-purple?style=for-the-badge" alt="Context Length"/>

<img src="https://img.shields.io/badge/Merge-C3SM--SDM-orange?style=for-the-badge" alt="Merge Algorithm"/>

<img src="https://img.shields.io/badge/Target-Agentic%20Coding%20%26%20Tools-red?style=for-the-badge" alt="Target"/>

</p>

Official GGUF quantizations for OliviaRossi/Qwen3.5-9B-C3SM-SDM-Agentic-Coder, a state-of-the-art 9B-parameter autonomous coding and multi-turn tool-calling model.

Synthesized using Curvature-Calibrated Spectral Consensus Merging with Sinusoidal Depth Modulation (C3SM-SDM), all multimodal and speculative MTP layers were surgically pruned to yield a pure, clean 32-block hybrid causal LM compatible with downstream llama.cpp runtimes and quantization schemes (Q4_K_M, Q5_K_M, Q8_0).

---

Model Lineage & Heritage

The model unifies a top-tier distillation anchor with four specialized domain-expert fine-tunes:

| Role | Repository | Domain Specialization |

| :--- | :--- | :--- |

| Anchor Base ($M_0$) | XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B | Reasoning distillation backbone, native instruction compliance, multi-turn state harness. |

| Donor 1 ($M_1$) | ornith-ai/Ornith-1.5-9B | Step-by-step logic, conversational reasoning, multi-turn format compliance. |

| Donor 2 ($M_2$) | Jackrong/Qwopus3.5-9B-Coder | Claude-Opus style agentic coding, deep tool-use schemas, recursive refactoring. |

| Donor 3 ($M_3$) | OrionLLM/OxCoder-9B | Complex algorithmic synthesis, long-context programming, MCP tool execution. |

| Donor 4 ($M_4$) | empero-ai/Qwen3.8-9B-Distill | Distilled Qwen3.8 foundation scale, high-capacity general world and syntax intelligence. |

| Chat Template | OliviaRossi/Improved-Chat-Template-for-Qwen-3.x | Dual-format agentic execution (JSON/XML), multi-tier <thought> reasoning CoT. |

---

Mathematical Architecture: C3SM-SDM

Standard weight merging algorithms (SLERP, TIES, DARE) apply static global scalars across the entire network, causing representation drift, activation saturation, and degraded token logits. C3SM-SDM merges across four continuous harmonic waveforms:

Normalized Layer Depth Coordinate: z = l / 31  (for layer l in [0..31])

 z = 0.0                      z = 0.5                      z = 1.0
 [Shallow Syntactic] ───> [Core Reasoning / Tools] ───> [Output Calibration]
  α(0) = 0.20                  α(0.5) = 0.85                 α(1) = 0.20
  p(0) = 0.50 (High DARE)      p(0.5) = 0.30 (Dense Delta)   p(1) = 0.50 (High DARE)
  r(0) = 32   (Low Rank)       r(0.5) = 96   (Rich Spectrum) r(1) = 32   (Low Rank)

1. Harmonic Task Amplitude Modulation: $\alpha(z)$

To prevent boundary degradation (token embedding corruption at layer 0 and logit distribution distortion at layer 31), the task arithmetic scale follows a harmonic envelope:

$$\alpha(z) = \alpha_{\min} + (\alpha_{\max} - \alpha_{\min}) \cdot \sin^2(\pi z), \quad \alpha \in [0.20, 0.85]$$

The core algorithmic manifold ($z \in [0.35, 0.75]$) receives the maximum fine-tuned coding delta, while syntactic boundaries remain securely anchored to the distilled base.

2. Phase-Shifted Domain Specialization: $\vec{\lambda}_k(z)$

Each donor model is routed according to its structural layer specialty:

$$w_k(z) = w_{k, \text{base}} \cdot \left[1.0 + A_k \cdot \sin(\pi z + \phi_k)\right], \quad \lambda_k(z) = \frac{w_k(z)}{\sum_j w_j(z)}$$

• Ornith-1.5 ($\phi = -0.20\pi$): Peaks in early reasoning layers ($z \approx 0.35$) to shape planning and CoT trajectories.
• Qwopus-Coder ($\phi = 0.00\pi$): Centers at mid-depth ($z \approx 0.50$) to govern agentic tool schemas and state transitions.
• OxCoder ($\phi = +0.25\pi$): Centers in deep reasoning blocks ($z \approx 0.65$) for heavy AST transformations and algorithm synthesis.
• Qwen3.8-Distill ($\phi = +0.05\pi$): Provides continuous foundational regularization across all depths.

3. Inverted Harmonic DARE Sparsification: $p(z)$

Parameter noise from fine-tuning collisions is most harmful to syntactic parsing at the boundaries. DARE dropout is inverted:

$$p(z) = p_{\max} - (p_{\max} - p_{\min}) \cdot \sin(\pi z), \quad p \in [0.30, 0.50]$$

4. Dynamic Spectral Subspace Denoising (rSVD): $r(z)$

Consensus deltas undergo randomized singular value decomposition to eliminate high-entropy orthogonal interference:

$$\Delta_{\text{consensus}} \approx \mathbf{U}_{r(z)} \mathbf{\Sigma}_{r(z)} \mathbf{V}_{r(z)}^T, \quad r(z) = \text{round}\left(32 + 64 \cdot \sin(\pi z)\right)$$

5. Frobenius Manifold Calibration

To ensure activation stability across GQA and Gated DeltaNet recurrent kernels, the merged weight tensor is projected back to the convex norm of its contributors:

$$\mathbf{W}^* = (\mathbf{W}_0 + \alpha(z) \Delta_{\text{spectral}}) \cdot \frac{(1-\beta)\|\mathbf{W}_0\|_F + \beta \sum \lambda_k \|\mathbf{W}_k\|_F}{\|\mathbf{W}_{\text{merged}}\|_F}$$

---

Quantization Matrix & Hardware Guidance

All quants are packaged as single-file GGUF binaries converted directly from the 32-block causal architecture:

| File Name | Precision | File Size | Recommended VRAM / RAM | Primary Use Case |

| :--- | :---: | :---: | :---: | :--- |

| Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Q5_K_M.gguf | 5-bit Medium | ~6.4 GiB | 12 GB VRAM (RTX 3060 12GB, RTX 4070) | Near-Lossless Precision: Superior syntax retention, complex multi-file refactoring, and strict schema adherence. |

---

Chat Template & Agent Prompting

The GGUF metadata includes the Olivia Rossi Improved Chat Template for Qwen 3.x. It natively parses structured tool calls, system-level declarations, and multi-tier Chain-of-Thought reasoning.

Tool Call Format (OpenAI / JSON Compatible)

<|im_start|>system
You are an expert autonomous coding agent.
<tools>
[
  {
    "name": "execute_bash",
    "description": "Execute a shell command in the workspace environment.",
    "parameters": {
      "type": "object",
      "properties": {
        "command": {"type": "string", "description": "The command to run."}
      },
      "required": ["command"]
    }
  }
]
</tools><|im_end|>
<|im_start|>user
Inspect the current directory and write a Python test runner script.<|im_end|>
<|im_start|>assistant
<thought>
The user wants to inspect the repository structure before implementing the test runner.
First step: list the directory contents using the bash tool.
</thought>
<tool_call>
{"name": "execute_bash", "arguments": {"command": "ls -la"}}
</tool_call><|im_end|>
<|im_start|>tool
<tool_response>
total 16
drwxr-xr-x 2 user user 4096 Sep 27 12:00 src
-rw-r--r-- 1 user user  420 Sep 27 12:00 pyproject.toml
</tool_response><|im_end|>
<|im_start|>assistant
<thought>
Directory contains `src/` and `pyproject.toml`. Now generate the pytest execution script.
</thought>
I have inspected the project directory. Let's create `run_tests.py`...<|im_end|>

---

Quickstart & Deployment

1. In Ollama

Create a Modelfile:

FROM ./Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Q5_K_M.gguf

PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER num_ctx 32768
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"

Build and launch:

ollama create qwen3.5-agentic -f Modelfile
ollama run qwen3.5-agentic

2. In llama-server (OpenAI-Compatible Local Endpoint)

To run as an autonomous agent backend for Aider, Cline, or Roo Code:

./llama-server \
    -m Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Q5_K_M.gguf \
    -c 32768 \
    -ngl 99 \
    --host 0.0.0.0 \
    --port 8080

Configure your IDE client:

• API Base: http://localhost:8080/v1
• Model ID: Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Q5_K_M
• API Key: sk-no-key-required

3. In llama-cli

./llama-cli \
    -m Qwen3.5-9B-C3SM-SDM-Agentic-Coder-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are a helpful coding assistant.<|im_end|>\n<|im_start|>user\nWrite a lock-free multi-producer single-consumer ring buffer in Rust.<|im_end|>\n<|im_start|>assistant\n<thought>\n" \
    -n 2048 \
    -c 8192 \
    -ngl 99

---

Architectural Verification

• Decoder Layers: 32 regular transformer blocks (blk.0 through blk.31).
• Linear Attention: Gated DeltaNet hybrid recurrence preserved with zero dropout perturbation.
• Multimodal Projectors: Stripped (visual., vision_tower., mmproj removed).
• MTP Heads: Speculative decoding layers eliminated for clean, zero-error GGUF load validation.
• Context Capacity: Up to 262,144 tokens supported via native RoPE scaling.

---

Citation & Acknowledgments

If you utilize this model or the C3SM-SDM merging formulation in your research or applications, please cite:

@misc{rossi2026c3smsdm,
  author = {Olivia Rossi},
  title = {Qwen3.5-9B-C3SM-SDM-Agentic-Coder: Curvature-Calibrated Spectral Consensus Merging with Sinusoidal Depth Modulation},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/OliviaRossi/Qwen3.5-9B-C3SM-SDM-Agentic-Coder-GGUF}}
}

Special gratitude to the authors of Xiaomi MiMo, Ornith AI, Jackrong (Qwopus), OrionLLM (OxCoder), and Empero AI for their open-weights contributions to the open-source coding ecosystem.