caid-technologies/parti-vision

Parti Vision

---

📄   Read the Whitepaper   |   💬   Join the Community

Turns a hardware idea — a sentence, a short brief, even a sketch — into a complete build blueprint.

Tell it what to build — "a USB-powered desk lamp with touch dimming" — optionally with a short document or a concept image, and it returns one structured blueprint plan: the parts, the wiring, ordered build steps, a costed sourcing table, and an appearance spec with a ready-to-use image-generation prompt. It's a standalone, all-in-one model (no adapter upon request).

Early research preview. For drafting and exploring ideas — not a replacement for real engineering, CAD, or safety review.

What it does

Give it a hardware idea and it returns, as one machine-readable JSON object:

• 📋 a parts list (electronics, printed parts, fasteners, with dimensions)
• 🔌 the wiring between parts, with a power budget and basic protection
• 🛠️ ordered build steps — fabrication → wiring → bring-up → assembly → testing
• 💲 costed sourcing whose line items add up
• 🎨 an appearance spec plus a ready-to-use image-generation prompt

Your app can parse, check, and build on the result directly.

Results

We test on requests the model has never seen during training. How often it produces a valid, well-structured blueprint:

On held-out requests	Stock Qwen3.5-9B	Parti-Vision (free)	Parti-Vision (guided)
Valid, well-structured blueprint	FAIL	67%	83%

On brand-new realistic requests, guided decoding reaches 97% valid blueprints (61% with free decoding); the stock base model manages FAIL on the same tests.

What's "guided decoding"? A standard serving option (guided_json in vLLM, "structured outputs" in most hosted APIs) that constrains the output to your blueprint format. The model still makes every design decision — the parts, the wiring, the steps, the costs — guided decoding just guarantees the shape.

What's improved

• FAIL → PASS The stock base model can't produce a valid blueprint; Parti-Vision does — 67% free, 83% guided, and 97% on realistic requests with guided decoding.
• Reads sketches, renders, and short briefs — not just text. Prompt + brief + render is strongest.
• General image understanding preserved — the visual pathway is left intact during fine-tuning.
• Answers in JSON directly — no reasoning preamble to strip.
• Cleaner than earlier iterations — no leaked reasoning, no markdown fences, no leading prose (all regression-checked).

What you can give it

• A plain-English request — one or two sentences.
• A short document — a brief or notes, pasted into the message.
• A concept image — a hand-drawn sketch or product render, as a vision input.

Any combination works; prompt + brief + render is strongest. The model reads text + images, so convert PDFs, LaTeX, CAD files, or spreadsheets to text or an image first.

Try it

The model answers in JSON directly — no reasoning preamble to strip.

from unsloth import FastVisionModel

REPO = "caid-technologies/parti-vision"
model, tok = FastVisionModel.from_pretrained(REPO, load_in_4bit=False)
FastVisionModel.for_inference(model)

SYSTEM_PROMPT = (
    "You design maker/electronics products. Given a request, reply with one JSON object "
    "describing the complete build — parts, wiring, build steps, sourcing, and appearance. "
    "Output only the JSON."
)
messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
    {"role": "user",   "content": [{"type": "text", "text": "Design a USB desk lamp with touch dimming."}]},
]
text = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tok(text=text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=13000, do_sample=False, repetition_penalty=1.1)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

💡 Blueprints are long: keep max_new_tokens high and the repetition penalty on. To include an image, add {"type": "image", "image": your_image} to the user content and pass images= to the tokenizer.

🚀 Serving it for real? Use vLLM with guided_json (or your engine's structured-outputs mode) constrained to your blueprint format — it takes valid-blueprint rates on unseen prompts from ~61% to ~97%.

Good to know

• English prompts, maker/electronics domain. Off-topic requests still get a blueprint, not a refusal.
• Outputs are drafts, not verified engineering — roughly half have at least one design slip (a wiring mistake, costs slightly off, a misordered step). Re-validate in your app, and review before you solder.
• Very long plans can get cut off at the token cap; contradictory or impossible requests can produce confidently wrong blueprints.

Learn more

• 📄 Technical whitepaper (PDF) — the model family, training approach, evaluation methodology, and the complete honest scorecard.
• 💬 Discord community — questions, builds, feedback.
• 🧩 Text-only or a smaller footprint? See caid-technologies/parti-base, the earlier 3B sibling.

Citation

@misc{parti_vision_base,
  title  = {Parti-Vision},
  author = {Caid Technologies},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/caid-technologies}}
}