Llama 4
Meta's open-weight Mixture-of-Experts model
What it is
Meta's open-weight family, the first to use a Mixture-of-Experts (MoE) architecture, and the most widely deployed open ecosystem in the world. Llama 4 Scout has the largest context window among open models (up to 10M tokens); Llama 4 Maverick beats previous flagship models on coding, reasoning, and images. Download it and run it locally without sending data to anyone, or use Groq/Together.ai for ultra-fast cloud inference.
How to use it, step by step
- 1
To run it locally install Ollama from ollama.com (Windows, Mac, Linux).
- 2
From a terminal run: ollama run llama4 — first launch downloads the model, then you talk to it offline.
- 3
Large MoE models want a powerful PC (GPU with lots of VRAM); on normal machines use the smaller variants.
- 4
For ultra-fast cloud inference without hardware, create an account on Groq (groq.com) or Together.ai and use their API.
- 5
Connect Llama to tools like Aider or LangChain by pointing them at the local Ollama endpoint.
💡Practical tips
- →Run locally and your data never leaves your PC: perfect for confidential documents.
- →Scout handles huge contexts (up to 10M tokens): ideal for analyzing entire archives.
- →Groq is impressive for speed: near-instant responses.
💰Pricing
The model is free (open weights). Locally it costs nothing; on Groq/Together you only pay for cloud inference.
The official site opens in a new tab.