Open Source Models

Introduction

Open-source LLMs (Llama, Mistral, Qwen, Gemma, DeepSeek) let you self-host, fine-tune freely and avoid per-token API costs. The trade-off: you manage GPUs, scaling and updates.

Beginner analogy: Open-source = owning a car. Hosted APIs = Uber. Both work; choose based on volume and control.

Understanding the topic

Core concepts to understand:

• Llama 3.1 (Meta) — flagship open weights, 8B–405B sizes.
• Mistral & Mixtral — efficient European models.
• Qwen 2.5 (Alibaba) — strong multilingual + coding.
• DeepSeek V3 / R1 — frontier reasoning, MIT license.
• Run via Ollama, vLLM, Together.ai, Groq, Fireworks.

Syntax reference

Visual workflow / architecture:

Hugging Face Hub
       │ download weights
       ▼
   Local GPU (Ollama)  or  Hosted (Groq, Together)
       │
       ▼
   Same OpenAI-compatible API
       │
       ▼
   Your application

Real-world use

Perplexity uses Llama for fast inference. Cursor ships with multiple model options. Many enterprises run open models on AWS Bedrock for data sovereignty.

Best practices

• Use hosted inference (Groq, Together) before self-hosting.
• Pick licenses carefully — Llama is mostly OK; some others are research-only.
• Benchmark on your tasks — open models can match GPT-4 on narrow domains.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Name three popular open-source LLMs.
• Q2. Why use open source over OpenAI?
• Q3. What's the trade-off?