A light weight Model that runs on a single GPU with over 80 TPS on a RTX 5060 (8G vram), Excellent for coding purposes.

Details

Updated 1 month ago

1 month ago

49689aef53ce · 2.5GB ·

model

archqwen3

parameters4.02B

quantizationQ4_K_M

2.5GB

system

You are Zentra — a next-generation AI built for people who don't settle for average. You are sharp

1.4kB

params

{ "num_ctx": 16384, "num_predict": 2048, "repeat_penalty": 1.1, "temperature": 0.7,

99B

VULCAN

Fast. Local. Private. Fine-tuned for developers who actually ship.

VULCAN is a fine-tuned 4B coding assistant built on Qwen3, trained on a curated mix of high-quality coding datasets (Magicoder + FineTome). It runs fully offline — no API keys, no rate limits, no data leaving your machine.

⚡ Quickstart

ollama run relational/VULCAN

🧠 What VULCAN Is Built For

VULCAN was fine-tuned specifically for real-world developer workflows, not academic benchmarks.

Domain	Capability
🐍 Python	Clean, production-ready scripts, APIs, automation
🌐 Web (HTML/CSS/JS)	Glassmorphism, animations, responsive layouts, no frameworks
🐛 Debugging	Trace errors, explain root cause, fix with context
⚙️ Backend	FastAPI, Express, REST architecture, auth flows
📐 System Design	Architecture decisions, trade-offs, scalability patterns
✍️ Technical Writing	Docs, READMEs, inline comments, changelogs
🔍 Code Review	Spot bugs, suggest improvements, refactor cleanly

🔧 Recommended Settings

# Larger context for long files and multi-file conversations
/set parameter num_ctx 16384

# Balanced creativity vs precision
/set parameter temperature 0.7

Thinking Mode (Qwen3 Feature)

Enable deep reasoning for hard problems:

/set think

Disable for fast, direct responses:

/set nothink

💡 Use /think for architecture decisions, debugging complex issues, or multi-step reasoning. Use /nothink for quick completions and boilerplate.

💬 Prompting Tips

Be specific — it makes a huge difference:

❌ Bad: make me a landing page

✅ Good: Create a dark-themed standalone HTML landing page with a hero section, glassmorphism cards, scroll-triggered animations using Intersection Observer, and a contact form with client-side validation. No external libraries.

❌ Bad: fix my code

✅ Good: This FastAPI endpoint returns a 422 error when I send a POST request with a nested JSON body. Here’s the route and the Pydantic model — what’s wrong?

Other tips: - Paste your full code when debugging — don’t summarize it - Mention your stack explicitly (e.g. “Python 3.11, FastAPI, PostgreSQL”) - Ask for explanations alongside fixes: “fix this and explain why it broke” - Use /think before complex architectural questions

📊 Benchmark Comparison

Compared against similarly-sized models available on Ollama:

Model	Params	Coding Focus	Context	Quant	Size
VULCAN (this)	4B	✅ Fine-tuned	16K	Q4_K_M	~2.5GB
qwen2.5-coder:3b	3B	✅ Yes	32K	Q4_K_M	~2GB
qwen2.5-coder:7b	7B	✅ Yes	128K	Q4_K_M	~4.7GB
phi4-mini	3.8B	⚠️ General	16K	Q4_K_M	~2.5GB
gemma3:4b	4B	⚠️ General	128K	Q4_K_M	~3.3GB
llama3.2:3b	3B	⚠️ General	128K	Q4_K_M	~2GB
deepseek-coder:6.7b	6.7B	✅ Yes	16K	Q4_K_M	~4.1GB

VULCAN sits in the sweet spot — coding-focused fine-tune, runs on modest hardware, fast inference.

🖥️ Performance Estimates

Hardware	Est. Speed	Notes
RTX 5060 8GB	~80–100 tok/s	Full GPU, very fast
RTX 4060 8GB	~65–85 tok/s	Full GPU
RTX 3060 12GB	~50–70 tok/s	Full GPU
RTX 3050 8GB	~35–50 tok/s	Full GPU
M2/M3 MacBook (16GB)	~40–60 tok/s	Metal acceleration
CPU only (16GB RAM)	~5–12 tok/s	Slow but works

📋 Hardware Requirements

	Minimum	Recommended
VRAM	4GB	6GB+
RAM	8GB	16GB
Disk	3GB free	5GB free
OS	Windows / macOS / Linux	Any

Runs on CPU if no GPU is available — just slower.

👤 Author

Made by Xlelords

Built from scratch — trained, quantized, and shipped locally. VULCAN is a personal project, not a corporate product.