Nebulos-Distill-Qwen3-0.6B

This is a lightweight reasoning model fine-tuned to perform efficient step-by-step logic. It was successfully distilled from the Qwen 3 architecture on consumer-grade hardware, proving that high-quality AI training is possible on a budget.

Model Details

Model Description

Nebulos-Distill is a compact 0.6B parameter model designed for high-speed local inference. It focuses on maintaining logical consistency and reasoning capabilities while requiring minimal VRAM.

• Developed by: Erik22TY 🐧
• Model type: Causal Language Model (Fine-tuned via LoRA)
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: unsloth/Qwen3-0.6B-bnb-4bit

Model Sources

• Repository: Erik22TY/Nebulos-Distill-Qwen3-0.6B

Uses

Direct Use

This model is intended for local deployment in reasoning-heavy tasks such as math word problems, logic puzzles, and concise text generation. It is highly recommended for mobile deployment or low-spec desktop environments.

Out-of-Scope Use

The model is 0.6B parameters; it should not be used for long-form creative writing or highly complex professional legal/medical advice as it may hallucinate due to its small size.

Training Details

Training Data

Fine-tuned using the AM-Qwen3-Distilled dataset, a high-quality collection of reasoning-oriented instructional data.

Training Procedure

Training Hyperparameters

• Training regime: fp16 (mixed precision) ⚡
• Optimizer: paged_adamw_8bit (to save VRAM)
• Gradient Accumulation Steps: 16
• Max Steps: 50
• Learning Rate: 2e-4

Speeds, Sizes, Times

• Hardware: NVIDIA GeForce GTX 1050 (3GB VRAM) 🏎️
• Training Time: ~1 hour and 15 minutes
• Final Loss: 0.9315 (Starting at 50 steps)
• Adapter Size: 4.60 MB

Environmental Impact

• Hardware Type: GTX 1050 Desktop
• Hours used: 1.25 hours
• Cloud Provider: N/A (Local Training on Linux Mint)

Technical Specifications

Compute Infrastructure

Hardware

• GPU: NVIDIA GTX 1050 (3.0 GB VRAM)
• OS: Linux Mint (Ubuntu-based)

Software

• Runtime: Ollama & PyTorch 2.5
• PEFT Library: LoRA (Rank 8)

How to Get Started with the Model

To run this model locally with Ollama, use:

ollama run hf.co/Erik22TY/Nebulos-Distill-Qwen3-0.6B:Q4_K_M