anshulVashist/Qwen3-8B-Houdini-VEX-v11

🪄 Qwen3-8B-Houdini-VEX-v11 (Grandmaster Edition)

<p align="center"> <img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200" alt="Unsloth"/> </p>

Qwen3-8B-Houdini-VEX-v11 is the flagship state-of-the-art neural code intelligence model fine-tuned specifically for SideFX Houdini VEX. Built on top of Qwen3-8B and trained with rigorous Compiler-in-the-Loop (CITL) validation, it is engineered to understand 3D vector spaces, quaternion orientations, SIMD CVEX parallelism, geometry attribute classes, and complex procedural graphics algorithms.

Developed by Anshul Vashist.

🌟 What's New in v11 (Grandmaster Edition)

• —🛡️ Compiler-in-the-Loop (CITL) Verification: Trained and audited against SideFX's native vcc compiler to eliminate syntax errors, deprecated functions, and signature hallucination.
• —🧠 Dual-Mode Architectural Reasoning:
• —⚡ Turbo Mode: Direct, compile-ready VEX snippets with zero conversational filler or markdown overhead.
• —🔬 Deep Reasoning Mode: Activates explicit <think>...</think> architectural analysis, solving 3D vector trigonometry, spatial hash lookups, and SIMD mechanics before emitting code.
• —📐 Rigorous Quaternion & Matrix Math: Native mastery of 3D rotations, p@orient = dihedral(...), qrotate, matrix3, and lookat transforms.
• —🌐 VDB & Volume Existence Guards: Automatic bounds checks, nametoprim guards, and volumetric gradient advection.
• —🔄 Negative Error Correction: Fine-tuned to avoid legacy GLSL habits (vec3 -> set), quaternion semantics (p@orient vs v@orient), and VEX syntax quirks (smooth vs smoothstep).

📦 Model Files in this Repository

💻 How to Use

Option 1: Live Inside SideFX Houdini (HDA & Live Viewport)

Use the official open-source `ai_attribwrangle.hda` digital asset with 1-click GUI installer and real-time viewport feedback: 👉 Houdini AI Wrangle GitHub: https://github.com/anshul-vashist/Houdini-Ai-Attribute-Wrangle

1. 1.Clone or download the GitHub repository.
2. 2.Run Setup_AI_Wrangle.exe or drop install_in_houdini.py into Houdini.
3. 3.Place Qwen3-8B-Houdini-VEX-v11-Q5_K_M.gguf into the models/ directory.
4. 4.Drop an AI Attribute Wrangle node into SOPs and generate procedural VEX directly on your active geometry!

Option 2: Standalone llama-server (Zero Dependencies)

Run the local inference engine with GPU offloading:

llama-server.exe \
  --model Qwen3-8B-Houdini-VEX-v11-Q5_K_M.gguf \
  --port 58421 \
  --host 127.0.0.1 \
  --ctx-size 4096 \
  --n-gpu-layers 36

Option 3: Hugging Face Transformers / Python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "anshulVashist/Qwen3-8B-Houdini-VEX-v11"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

prompt = "Context: Point wrangle\nTask: Displace geometry along @N with 3D curl noise"
messages = [
    {"role": "system", "content": "You are an expert Houdini VEX compiler. Output ONLY pure VEX snippet code. No markdown, no explanations."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))

⚡ Prompting Format

Turbo Fast Mode (Instant VEX):

<|im_start|>system
You are an expert Houdini VEX compiler. Output ONLY pure VEX snippet code. No markdown, no backticks, no explanations, no HTML tags.<|im_end|>
<|im_start|>user
Context: Point wrangle
Task: Displace points along normal with 3D curl noise
Write pure Houdini VEX code:<|im_end|>
<|im_start|>assistant

Deep Reasoning Mode:

<|im_start|>system
You are Houdini VEX Copilot, an expert AI programming assistant fine-tuned and created by Anshul Vashist.
When presented with a task, FIRST think through the problem in an explicit <think>...</think> block:
  1. Analyze input geometry, vector spaces, and physical/mathematical equations.
  2. Outline algorithmic steps and procedural logic (loops, spatial lookups, matrix transforms).
AFTER the </think> tag, output ONLY the pure, verified Houdini VEX code.<|im_end|>
<|im_start|>user
Context: Detail wrangle
Task: Generate a Fibonacci spiral distribution of 200 points on a unit sphere using golden ratio angle
Write pure Houdini VEX code:<|im_end|>
<|im_start|>assistant
<think>

📜 License

Apache-2.0. Free for commercial and personal procedural graphics production.