Meta Llama 4: Everything You Need to Know
Llama 4 Scout and Maverick are open-weight, natively multimodal MoE models. Scout advertises a 10M-token context. Here is how to run or chat with them.
Earlier snapshot: Meta Llama 3.
The current open-weight family is Llama 4.
Meta’s Llama 4 announcement introduced Scout and Maverick: natively multimodal mixture-of-experts models. Scout is the efficiency SKU (fits on a single H100 with Int4) and is documented with a 10 million token context window. Maverick is the fast, cheaper text-and-image model. A larger teacher, Behemoth, was previewed as still training.
Weights are on llama.com and Hugging Face. Our matrix page is Llama; the consumer chat surface is Meta AI.
Open weights, still the point
Unlike ChatGPT or Claude, you can download the weights, run them on your hardware, and fine-tune without an API contract. That is why Llama shows up in local runtimes, Groq, and every on-prem stack.
Our context window row for Llama is 128k for the product we tracked as a chat tool — Scout’s 10M figure is the model card for that SKU, not a promise that meta.ai’s chat box will take a 10M paste.
Where to try it without a GPU
- meta.ai — Meta’s official chat.
- WhatsApp, Messenger, Instagram Direct — Meta wired Llama 4 into those surfaces on launch day.
- Llama API — playground plus an OpenAI-compatible SDK.
- Groq / Cerebras — Meta’s LlamaCon post listed them as fast-inference partners on the Llama API.
Why Meta still gives it away
Same strategy as Llama 3: commoditize the model layer so Meta is not locked into a competitor’s API, and let the ecosystem fine-tune. If you need the weights on-prem, start at /tools/llama. If you just want a free chat box, use /tools/metaai.