zveloxy

Veloxy-Titan-32B-Vision

README

License: apache-2.0

📌 1. Overview

Veloxy-Titan-32B-Vision is the premier multimodal visual intelligence and deep reasoning flagship of the Veloxy AI family. It fuses a high-resolution 27-layer Vision Transformer (ViT) tower (hidden size: 1,152, patch size: 16x16) directly with the 32B GRPO-refined logical reasoning backbone into a unified, native bfloat16 weight representation.

The model parses visual inputs, UI/UX mockups, architectural schematics, printed circuit board (PCB) diagrams, complex multi-column tables, and high-DPI scans pixel-by-pixel, directly generating production-ready code, structured JSON, or in-depth technical analysis.

⚙️ 2. Architectural Specifications

Table with columns: Component / Parameter, Value, Technical Details
Component / Parameter	Value	Technical Details
Vision Tower	27-Layer ViT (Vision Transformer)	Hidden Size: 1,152, Patch Size: 16x16, 5,120 Cross-Attention Projection
Language Backbone	32B Class (27.2B Active)	Hybrid Linear-Attention + Full Multi-Head Attention Blocks
Total Weight Tensors	1,184 Tensors	End-to-end multimodal weight integration
Native Precision (Dtype)	`torch.bfloat16`	12 Safetensors shards (~54 GB total)
Image Resolution Handling	Dynamic Multi-Patch ViT Encoding	Arbitrary aspect ratio, native high-DPI support
Context Window	131,072 Tokens (128K+)	Interleaved high-resolution visual tokens and long text
Temporal Anchor	2026	Modern UI standards (React 19, Next.js 15, Tailwind CSS v4)
License	Apache 2.0	Permissive open source for commercial and research use

🔬 3. Live Visual Verification Benchmark (NVIDIA H100 NVL)

The model was tested live on an enterprise NVIDIA H100 NVL (93.1 GB HBM3) node using a synthetic Dark Mode Telemetry Dashboard UI mockup:

🎯 Telemetry Extraction & Accuracy:

Visual Pixel Perception:
- Exact background detection: #0f172a (Slate-900).
- Sky-blue P99 latency metric card (P99 LATENCY: 1.24ms) accurately parsed.
- Emerald cluster health badge (CLUSTER HEALTH: OPTIMAL 99.99%).
- Purple node counter (ACTIVE NODES: 32).
- Interactive action button (DEPLOY TO PRODUCTION) with hover states.
Code Synthesis: The entire visual hierarchy was converted zero-shot into a Next.js 15, React 19, and Tailwind CSS component without syntax errors.
Throughput: Synthesized 650 tokens of clean code in native bfloat16 at 10.09 tokens/second.

🌟 4. Core Multimodal Capabilities

🎨 1. Pixel-to-Code & UI/UX Reverse Engineering

Parses Figma exports, mobile screenshots, or web wireframes into clean, accessible code:

React 19 / Next.js 15 (Tailwind CSS): Semantic, responsive components with Tailwind typography and flexbox/grid layout.
Flutter & SwiftUI: Native declarative UI code for iOS and Android.
HTML5 / CSS: Clean, zero-framework semantic web layouts.

📐 2. Technical Schematics & Circuit Diagram Analysis

Cloud Architecture Schematics: Translates AWS, Azure, GCP, and Kubernetes architecture diagrams directly into Terraform / OpenTofu HCL code.
Electronics & PCB Schematics: Recognizes component designations (resistors, IC pinouts, capacitors) and conducts circuit sanity checks.
Database ERD & UML: Extracts entity-relationship diagrams into Prisma, Drizzle, or raw SQL migrations.

📄 3. Document AI & Complex Table Intelligence

Multi-Column Financial Statements: Accurately extracts balance sheets, footnotes, and merged table cells into structured JSON or Markdown.
Scanned Contracts & Legal Forms: Transcribes stamped, signed, or multi-column documents with layout preservation.

🛡️ 4. Visual Threat Modeling & Network Security

Audits network topology maps, firewall rule diagrams, and cloud IAM graphs:

Detects public exposure paths, lateral movement vectors, and DMZ misconfigurations.

🚀 5. Quickstart & Serving Guide

A. Python / Transformers (Native bfloat16)

python

import torchfrom transformers import AutoProcessor, AutoModelForImageTextToTextfrom PIL import Image model_id = "zveloxy/Veloxy-Titan-32B-Vision" processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForImageTextToText.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto",    trust_remote_code=True) # Load target imageimage = Image.open("dashboard_mockup.png") messages = [    {        "role": "user",        "content": [            {"type": "image"},            {                "type": "text",                 "text": "Analyze all UI components, color palettes, and typography in this screenshot. Generate a production-ready Next.js 15 and Tailwind CSS component."            }        ]    }] prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[prompt], images=[image], return_tensors="pt").to("cuda") outputs = model.generate(    **inputs,    max_new_tokens=1500,    temperature=0.4,    top_p=0.9) print(processor.decode(outputs[0], skip_special_tokens=True))

B. High-Throughput Multimodal Serving with vLLM

bash

vllm serve zveloxy/Veloxy-Titan-32B-Vision \    --dtype bfloat16 \    --max-model-len 16384 \    --gpu-memory-utilization 0.95 \    --trust-remote-code \    --port 8000

C. Local Ollama & llama.cpp (GGUF + mmproj)

Thanks to the open-source community release by mradermacher/Veloxy-Titan-32B-Vision-GGUF, both the language weights and the multimodal projector (mmproj) are readily available:

dockerfile

# ModelfileFROM ./Veloxy-Titan-32B-Vision.Q4_K_M.ggufMMPROJ ./Veloxy-Titan-32B-Vision.mmproj-f16.gguf PARAMETER temperature 0.4PARAMETER top_p 0.9

bash

ollama create veloxy-titan-vision -f Modelfileollama run veloxy-titan-vision "Describe this architecture" diagram.png

🌐 6. Veloxy AI Model Ecosystem

Table with columns: Model, Format / Parameters, Focus & Target Hardware, Link
Model	Format / Parameters	Focus & Target Hardware	Link
Veloxy-Titan-32B-Vision 👁️	32B + 27-Layer ViT (BF16)	Multimodal Vision & Reasoning, UI-to-Code, Schematics	HF Repo
Veloxy-Titan-32B ⭐	32B Safetensors (BF16)	Flagship Deep Reasoning, 1M Context, H100 / Datacenter	HF Repo
Veloxy-Flash-15B (Faz 8) ⚡

📄 7. License

Veloxy-Titan-32B-Vision is licensed under the Apache License 2.0.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

zveloxy

Model Tree

Base

Qwen/Qwen3.8-27B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

📌 1. Overview

⚙️ 2. Architectural Specifications

Table with columns: Component / Parameter, Value, Technical Details
Component / Parameter	Value	Technical Details
Vision Tower	27-Layer ViT (Vision Transformer)	Hidden Size: 1,152, Patch Size: 16x16, 5,120 Cross-Attention Projection
Language Backbone	32B Class (27.2B Active)	Hybrid Linear-Attention + Full Multi-Head Attention Blocks
Total Weight Tensors	1,184 Tensors	End-to-end multimodal weight integration
Native Precision (Dtype)	`torch.bfloat16`	12 Safetensors shards (~54 GB total)
Image Resolution Handling	Dynamic Multi-Patch ViT Encoding	Arbitrary aspect ratio, native high-DPI support
Context Window	131,072 Tokens (128K+)	Interleaved high-resolution visual tokens and long text
Temporal Anchor	2026	Modern UI standards (React 19, Next.js 15, Tailwind CSS v4)
License	Apache 2.0	Permissive open source for commercial and research use

🔬 3. Live Visual Verification Benchmark (NVIDIA H100 NVL)

The model was tested live on an enterprise NVIDIA H100 NVL (93.1 GB HBM3) node using a synthetic Dark Mode Telemetry Dashboard UI mockup:

🎯 Telemetry Extraction & Accuracy:

Visual Pixel Perception:
- Exact background detection: #0f172a (Slate-900).
- Sky-blue P99 latency metric card (P99 LATENCY: 1.24ms) accurately parsed.
- Emerald cluster health badge (CLUSTER HEALTH: OPTIMAL 99.99%).
- Purple node counter (ACTIVE NODES: 32).
- Interactive action button (DEPLOY TO PRODUCTION) with hover states.
Code Synthesis: The entire visual hierarchy was converted zero-shot into a Next.js 15, React 19, and Tailwind CSS component without syntax errors.
Throughput: Synthesized 650 tokens of clean code in native bfloat16 at 10.09 tokens/second.

🌟 4. Core Multimodal Capabilities

🎨 1. Pixel-to-Code & UI/UX Reverse Engineering

Parses Figma exports, mobile screenshots, or web wireframes into clean, accessible code:

React 19 / Next.js 15 (Tailwind CSS): Semantic, responsive components with Tailwind typography and flexbox/grid layout.
Flutter & SwiftUI: Native declarative UI code for iOS and Android.
HTML5 / CSS: Clean, zero-framework semantic web layouts.

📐 2. Technical Schematics & Circuit Diagram Analysis

Cloud Architecture Schematics: Translates AWS, Azure, GCP, and Kubernetes architecture diagrams directly into Terraform / OpenTofu HCL code.
Electronics & PCB Schematics: Recognizes component designations (resistors, IC pinouts, capacitors) and conducts circuit sanity checks.
Database ERD & UML: Extracts entity-relationship diagrams into Prisma, Drizzle, or raw SQL migrations.

📄 3. Document AI & Complex Table Intelligence

Multi-Column Financial Statements: Accurately extracts balance sheets, footnotes, and merged table cells into structured JSON or Markdown.
Scanned Contracts & Legal Forms: Transcribes stamped, signed, or multi-column documents with layout preservation.

🛡️ 4. Visual Threat Modeling & Network Security

Audits network topology maps, firewall rule diagrams, and cloud IAM graphs:

Detects public exposure paths, lateral movement vectors, and DMZ misconfigurations.

🚀 5. Quickstart & Serving Guide

A. Python / Transformers (Native bfloat16)

python

import torchfrom transformers import AutoProcessor, AutoModelForImageTextToTextfrom PIL import Image model_id = "zveloxy/Veloxy-Titan-32B-Vision" processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForImageTextToText.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto",    trust_remote_code=True) # Load target imageimage = Image.open("dashboard_mockup.png") messages = [    {        "role": "user",        "content": [            {"type": "image"},            {                "type": "text",                 "text": "Analyze all UI components, color palettes, and typography in this screenshot. Generate a production-ready Next.js 15 and Tailwind CSS component."            }        ]    }] prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = processor(text=[prompt], images=[image], return_tensors="pt").to("cuda") outputs = model.generate(    **inputs,    max_new_tokens=1500,    temperature=0.4,    top_p=0.9) print(processor.decode(outputs[0], skip_special_tokens=True))

B. High-Throughput Multimodal Serving with vLLM

bash

vllm serve zveloxy/Veloxy-Titan-32B-Vision \    --dtype bfloat16 \    --max-model-len 16384 \    --gpu-memory-utilization 0.95 \    --trust-remote-code \    --port 8000

C. Local Ollama & llama.cpp (GGUF + mmproj)

Thanks to the open-source community release by mradermacher/Veloxy-Titan-32B-Vision-GGUF, both the language weights and the multimodal projector (mmproj) are readily available:

dockerfile

# ModelfileFROM ./Veloxy-Titan-32B-Vision.Q4_K_M.ggufMMPROJ ./Veloxy-Titan-32B-Vision.mmproj-f16.gguf PARAMETER temperature 0.4PARAMETER top_p 0.9

# ModelfileFROM ./Veloxy-Titan-32B-Vision.Q4_K_M.ggufMMPROJ ./Veloxy-Titan-32B-Vision.mmproj-f16.gguf PARAMETER temperature 0.4PARAMETER top_p 0.9

bash

ollama create veloxy-titan-vision -f Modelfileollama run veloxy-titan-vision "Describe this architecture" diagram.png

🌐 6. Veloxy AI Model Ecosystem

Table with columns: Model, Format / Parameters, Focus & Target Hardware, Link
Model	Format / Parameters	Focus & Target Hardware	Link
Veloxy-Titan-32B-Vision 👁️	32B + 27-Layer ViT (BF16)	Multimodal Vision & Reasoning, UI-to-Code, Schematics	HF Repo
Veloxy-Titan-32B ⭐	32B Safetensors (BF16)	Flagship Deep Reasoning, 1M Context, H100 / Datacenter	HF Repo
Veloxy-Flash-15B (Faz 8) ⚡

📄 7. License

Veloxy-Titan-32B-Vision is licensed under the Apache License 2.0.