Open-Source Models — Llama, Mistral, DeepSeek, Qwen

The open-weight ecosystem — self-hosted AI with Llama 4, Mistral, DeepSeek, Qwen

Open-Source Models

Major Open Model Families

Llama 4 (Meta, Feb 2026)

Scout (109B MoE) and Maverick (400B MoE)

1M context window — industry-leading for open models

Free for research and commercial use

DeepSeek V3 & R1

V3: 685B params, trained for $5.6M

R1: Open-source reasoning model (like o3, but open)

API: $0.27/M input (V3), $0.14/M (R1)

Mistral AI

Large 2 (July 2024): 128K, strong multilingual

Small 3 (Jan 2026): 24B, Apache 2.0, runs on consumer GPUs

Qwen 3 (Alibaba, Apr 2026)

235B MoE flagship, Apache 2.0

Strong multilingual performance

When to Choose Open-Source

Factor	Open-Source	Proprietary
Cost at scale	Free self-hosting	Pay per token
Control	Full customization	Limited to API
Data privacy	Stays on your hardware	Sent to provider
Ease	Needs infrastructure	Just an API call

Your Turn!

Compare VRAM requirements:

python

models = [
    ("Llama 4 Scout", "~60GB (4-bit)"),
    ("Mistral Small 3", "~16GB (4-bit)"),
    ("DeepSeek V3", "~340GB (8-bit)"),
    ("Qwen 3 (235B)", "~130GB (8-bit)"),
]
for name, vram in models:
    print(f"  {name:20s} | VRAM: {vram}")

✏️ Code Editor

Loading Python...

📤 Output

Write your solution and click "Run Code" to test it!

← Evaluation & Benchmarks 🎉 Course Complete!