license: apache-2.0 datasets: - Magpie-Align/Magpie-Pro-MT-300K-v0.1 - Magpie-Align/Magpie-Qwen2.5-Coder-Pro-300K-v0.1 - Magpie-Align/Magpie-Llama-3.3-Pro-500K-Filtered language: - de - en base_model: - Qwen/Qwen3.5-4B pipeline_tag: text-generation library_name: adapter-transformers tags: - efficient - qwen - qwen3.5 - gguf - ollama - instruction-finetuning - nomi - lazyloopstudio - unsloth - nomi2.0

Nomi 2.0

Introduction

Nomi-2.0 is a refined mid-range Large Language Model based on the Qwen-3.5-4B architecture. It was specifically developed to outperform standard 4B models in structured reporting, Markdown formatting, and Python coding, making it an ideal assistant for local deployment on consumer hardware.

In this training, we aimed to improve Nomi's reasoning. It's base, Qwen 3.5 4B was create at most thinks, but overthought most of the requests und looped a lot. Nomi 2.0 fixes that. Even at more complicated prompts it only thinks for about 10 sec (~500 Reasoning tokens) It is our second model in the Nomi series.

🌟 Key Features & Improvements

Architecture: Qwen-3.5-4B (runs on 8 GB VRAM GPUs like the RTX 4060).
Multilingual Support: Can understand and generate text in German and English.
Efficiency: High-speed inference (~60 tokens/sec) at full precision (83.7 tokens/sec with 8-bit quantization), with a ~0.05–0.3 s delay to the first token.

🧠 Training Details

The goal of Nomi-2.0 is to create a "bridge" model that feels as smart as a 7B model but runs with the speed and efficiency of a 4B model.

Base Model: Qwen/Qwen3.5-4B
Fine-tuning: SFT (Supervised Fine-Tuning).
Training Tool: Unsloth (for 4-bit optimized training).

⚠️ Limitations

As a 4B parameter model, Nomi-2.0 is not a replacement for Claude or larger models when it comes to deep world knowledge or complex mathematical reasoning. It is a specialized tool for speed, local privacy, and high-quality document structure.