Llama 4 (presentation on https://www.llama.com) is Meta’s new language model available in two versions, Scout (demo: <a href="https://huggingface.co/spaces/openfree/Llama-4-Scout-17B-Research”>https://huggingface.co/spaces/openfree/Llama-4-Scout-17B-Research) and Maverick (demo: <a href="https://huggingface.co/spaces/openfree/Llama-4-Maverick-17B-Research”>https://huggingface.co/spaces/openfree/Llama-4-Maverick-17B-Research). They have 17 billion parameters and accept up to 10 million tokens.
A third version, Behemoth, is in preparation.

These two models use the technique of coordinated “experts” (16 for Scout and 128 for Maverick): they divide the instruction received into several tasks and each “expert” will solve one or more of these tasks in connection with a part of the database rather than querying the entire database.

Free, open-source and unlimited on both demo spaces.