Llama 4

The Llama 4 collection of models are natively multimodal AI models that enable text and multimodal experiences. These models leverage a mixture-of-experts architecture to offer industry-leading performance in text and image understanding. Can't wait to try this out. We're experimenting with running models on-device for our product (desktop app) but haven't been able to get great results yet for the average laptop. Looking forward to see the reality of inference speeds for these models.

Llama 4 Scout:

•⁠ 17B x 16 experts

•⁠ Natively multi-modal

•⁠ 10M token context length

•⁠ Runs on a single GPU

•⁠ Highest performing small model

Llama 4 Maverick:

•⁠ 17B x 128 experts

•⁠ Natively multi-modal

•⁠ Beats GPT-4o and Gemini Flash 2

•⁠ Smaller and more efficient than DeepSeek, but still comparable on text, plus also multi-modal

•⁠ Runs on a single host

Llama 4 Behemoth:

•⁠ ⁠2+ trillion parameters

•⁠ ⁠Highest performing base model

•⁠ Still training!