🪄 Qwen3-8B-Houdini-VEX-v11 (Grandmaster Edition)
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
vcccompiler 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,
nametoprimguards, and volumetric gradient advection. - 🔄 Negative Error Correction: Fine-tuned to avoid legacy GLSL habits (
vec3->set), quaternion semantics (p@orientvsv@orient), and VEX syntax quirks (smoothvssmoothstep).
📦 Model Files in this Repository
| File | Size | Description |
|---|---|---|
Qwen3-8B-Houdini-VEX-v11-Q5_K_M.gguf | 5.85 GB | Recommended balanced GGUF for 8GB–12GB VRAM GPUs (llama.cpp / Ollama) |
Qwen3-8B-Houdini-VEX-v11-Q8_0.gguf | 8.71 GB | Near-lossless high-precision GGUF for 16GB+ VRAM workstations |
lora_adapter/ | ~700 MB | PEFT safetensors adapter weights (Rank 64, Alpha 128) for transformers/unsloth |
model-*.safetensors | 15.6 GB | Complete 16-bit unquantized full precision weights |
💻 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
- Clone or download the GitHub repository.
- Run
Setup_AI_Wrangle.exeor dropinstall_in_houdini.pyinto Houdini. - Place
Qwen3-8B-Houdini-VEX-v11-Q5_K_M.ggufinto themodels/directory. - 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.