caid-technologies-parti-vision-GGUF

caid-technologies-parti-vision-GGUF

Parti-Vision Base is a Qwen3.5-9B fine-tune by caid-technologies that turns a hardware idea — a sentence, a brief, or a sketch — into a complete build blueprint, returning a single structured JSON object containing a parts list, pin-level wiring, ordered build steps, a costed sourcing table, and an appearance spec with a ready-to-use image-generation prompt. It's a standalone, adapter-free merged 16-bit model (a LoRA-adapter version is also available) trained via Unsloth on 150 synthetic schema- and semantic-gate-validated records across four input variants (prompt-only, plus brief, plus sketch, plus brief+render), accepting plain-English requests, pasted documents, or concept images/sketches as vision input, and defaulting to direct JSON output with enable_thinking=False (though a reasoning trace can be enabled). While the stock base Qwen3.5-9B scores 0% on blueprint validity, Parti-Vision achieves ~67% schema-valid in-distribution and ~61% on unseen realistic prompts with free decoding — jumping to ~83% and ~97% respectively when served with vLLM's guided_json constrained decoding — though it remains an early research preview where roughly half of outputs contain at least one design slip (wrong pin nets, cost errors, misordered steps), making it a tool for rapid ideation and drafting rather than a substitute for real engineering, CAD, or safety review.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp