Open Source Models — Which One to Choose?

Episode 4 20 min

So far we have mostly worked with closed-source models like ChatGPT and Claude. But there is a vast world of open-source models that you can download for free, run on your own system, and even modify. Let us explore them.

What Does Open Source Mean for AI?

When we say a model is open-source, it means the model weights — the numbers that form the brain of the model — are publicly released. You can download and run it on your own computer. No internet needed, no monthly payments, and most importantly: your data stays with you.

The Llama Family — Meta’s Giant

Llama (Large Language Model Meta AI) by Meta is probably the most well-known open-source model family. From Llama 1 (Feb 2023) through Llama 4 (2025) with its MoE architecture, each version brought significant improvements. Strengths include a huge community, excellent tool compatibility, relatively open licensing, and sizes from 1B to 405B parameters.

The Qwen Family — Great for Non-English Languages

Qwen by Alibaba Cloud is a surprise performer, especially for non-English languages. Trained on 20+ languages with strong multilingual support, Qwen 3 includes Thinking Mode for step-by-step reasoning. Available from 0.6B to 235B parameters, the 8B and 14B versions offer the best quality-to-hardware ratio.

DeepSeek — Reasoning Power

DeepSeek focused on better thinking rather than just bigger models. DeepSeek-R1 is a reasoning model that thinks step-by-step before answering (Chain-of-Thought). It uses Mixture of Experts (MoE) architecture — like having a team of specialists instead of one generalist.

How to Choose the Right Model

By Use Case

  • General chat: Qwen 3 (8B or 14B)
  • Coding: Qwen-Coder or DeepSeek-Coder
  • Reasoning and math: DeepSeek-R1 or Qwen 3 with Thinking Mode
  • General English work: Llama 3 (8B or 70B)

Quantization Rule of Thumb

A 14B model with Q4 quantization usually outperforms a 7B model at FP16. If you have limited GPU memory, choose a larger model with more quantization.

Tools for Running Open Source Models

  • Ollama — Simplest way. One command to download and run.
  • LM Studio — GUI-based, beginner friendly.
  • vLLM — For production serving with API compatibility.

Summary

  1. Open-source vs closed-source trade-offs
  2. Three main families: Llama (popular), Qwen (great multilingual), DeepSeek (strong reasoning)
  3. Choose based on size, use case, and hardware
  4. Quantization is key for running big models on modest hardware