⚡ Qwen3-8B-Houdini-VEX-v30 (Master Edition)
Qwen3-8B-Houdini-VEX-v30 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 fine-tuned on Google Colab Cloud Tesla T4 GPUs with rigorous Compiler-in-the-Loop (CITL) verification, it is engineered for procedural geometry, dynamic parameter synthesis, quaternion orientations, SIMD CVEX parallelism, geometry attribute classes, and complex procedural graphics algorithms.
Developed by Anshul Vashist.
🌟 What's New in v30 (Master Edition)
- 🛡️ Autonomous Compiler-in-the-Loop (CITL) Verification: Trained and audited against SideFX's native
vcccompiler and physical runtime (hython.exeHoudini 21) 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 (<2s generation).
- 🔬 Deep Reasoning Mode: Conducts 3D vector space analysis, trigonometric calculations, and algorithmic planning inside an explicit
<think>...</think>block before emitting pristine VEX code.
- 🎛️ Automatic Dynamic Parameter Synthesis: Automatically writes code bound to Houdini channel evaluation functions (
chf,chi,chv,chs,chramp,chp), enabling instant UI slider creation. - 📦 6,167 Live Verified Training Records: Complete catalog coverage across SOPs, DOPs, POPs, OpenVDB, CHOPs, COPs, LOPs/USD, KineFX, and Position-Based Dynamics (PBD).
📊 Training Specifications
| Parameter | Specification |
|---|---|
| Base Architecture | unsloth/Qwen3-8B (Qwen 3 Causal LM) |
| LoRA Rank / Alpha | Rank 64, Alpha 128 across all 7 linear projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj) |
| Training Hardware | Cloud Tesla T4 GPU (Google Colab) |
| Training Steps | 154 steps (1 full epoch) |
| Final Loss | Train Loss: 0.3248 |
| Quantization | Standalone Merged Q5_K_M (~5.45 GB, optimal for 6–8 GB VRAM) |
🚀 Quickstart & Deployment
1. Embedded Local Inference (llama.cpp)
Run the standalone merged GGUF with llama-server on port 58421:
llama-server.exe --model Qwen3-8B-Houdini-VEX-v30-Q5_K_M.gguf --port 58421 --host 127.0.0.1 --ctx-size 4096 --n-gpu-layers 36
2. Ollama Deployment
Use the included Modelfile:
ollama create qwen3-vex-v30 -f Modelfile
ollama run qwen3-vex-v30
3. Native Houdini Integration (ai_attribwrangle.hda)
Drop ai_attribwrangle.hda into your Houdini SOP network. The HDA automatically connects to port 58421 and uses v30 Master with zero setup.
⚖️ License & Attribution
Licensed under the Apache 2.0 License. Created by Anshul Vashist for procedural visual effects artists and technical directors worldwide.