Jally AI

by Lazy Loop Studio

Small, fast & indie‑built language models for real people, not just benchmarks.

🌱 Our Models

Nomi 1.0 · 3B

Nomi-1.0-3B

Base model with core instruction tuning and strong Markdown + German/English support. Perfect for local deployment.

  • GGUF & .safetensors formats
  • Runs on consumer hardware
  • Apache 2.0 license
Visit on Hugging Face
Nomi 1.1 · 3B

Nomi-1.1-GGUF

Refined behavior, better formatting, improved multilingual abilities, optimized for local inference.

  • Fast ~80 tokens/sec
  • 0.05–0.3s first token
  • Enhanced German/English
Visit on Hugging Face
Coming Soon

Nomi-2 · 4B

Based on Qwen 3.5 4B with RASV reasoning style. Out now - Check it out!

  • Reasoning model (Restatement, Approach, Step-by-step, Verification)
  • ~500 reasoning tokens on complex prompts
  • Strong multilingual support
Visit on Hugging Face

🧩 Indie Spirit

Nomi AI is a small, independent project driven by a passion for small but strong models, not by a corporate budget. We experiment with modern base models, fine‑tune them on carefully selected datasets, and try to squeeze as much "brain" as possible out of just 3B parameters.

🔧 What We Build

🧠 Compact Models

3B models that feel like larger models — Nomi-1.0 and Nomi-1.1 deliver surprising capability on consumer hardware.

📝 Structured Output

Assistants focused on clean Markdown, tables, H1/H2 headers, and developer-friendly formatting.

🌍 European Focus

Strong German/English support for European developer setups, with multilingual improvements in 1.1.

🗒️ Our Roadmap

1 Nomi 1.0

Basic instruction tuning

2 Nomi 1.1

Improved behavior + code + math

3 Nomi 2.0

Coding, math, reasoning

4 Agentic

Tool-using tasks