Two released models and one preview
Meta released Llama 4 Scout and Llama 4 Maverick on April 5, 2025. They were the first Llama models designed as natively multimodal systems and the first in the family to use a mixture-of-experts architecture. Meta also previewed a much larger teacher model called Behemoth, but the downloadable release covered Scout and Maverick rather than Behemoth.
Both released models activate about 17 billion parameters for each token. Scout contains 16 routed experts and about 109 billion total parameters, while Maverick uses 128 routed experts and about 400 billion total parameters. A router directs each token through only part of the network. This reduces active computation compared with a dense model of the same total size, though all weights still need to be stored and moved efficiently.
Native multimodality and long context
Llama 4 was trained with text and visual information in a shared model backbone. That early-fusion approach lets the model reason over combinations of text and images instead of attaching a separate vision component only after language training. The released instruction models generate text, not new images, but can use images as input for description, analysis, and question answering.
Meta advertised a ten-million-token context window for Scout and a one-million-token window for Maverick. A maximum window describes what the architecture and serving stack can accept; it does not promise perfect retrieval across every position. Long-context quality must be measured on the actual document mix and task. Memory capacity, attention implementation, quantization, and serving software also determine whether a local deployment can approach the published limit.
Meta said Scout could fit on one H100 GPU when quantized to Int4 and that Maverick could run on a single H100 host. Those are specific configurations, not ordinary consumer-hardware requirements. The official repository notes that full-precision Llama 4 inference requires multiple GPUs. Organizations should calculate memory, throughput, energy, and operational cost for their chosen precision and workload.
What “open” means here
Scout and Maverick were released as open-weight models: approved users could download the parameters and Meta supplied inference code, a model card, an acceptable-use policy, and a community license. This supports private hosting, fine-tuning, research, and integration without sending every request to Meta’s hosted service.
The release is not open source in the strict sense used for software with an OSI-approved license. The Llama 4 Community License includes use conditions and special terms for services above a stated monthly-active-user threshold. Meta also did not provide the complete training corpus or a fully reproducible training pipeline. “Open weight” is therefore the precise description.
Benchmarks and important caveats
Meta reported that Scout and Maverick performed strongly against other models in their classes across multimodal, coding, reasoning, and multilingual tests. It also described an experimental chat version of Maverick used in an arena evaluation. Readers should avoid treating that experimental score as a measurement of every downloadable checkpoint. Developer-reported benchmark tables mix model variants, prompts, inference settings, and sometimes different dates.
The model card documents further limitations. Llama 4 can generate false information, unsafe suggestions, biased language, and errors when interpreting images. Its support spans many languages, but performance is not uniform. Deployers need application-specific evaluation, content safeguards, monitoring, and a clear way for users to challenge an output.
Why it matters
Llama 4 made three frontier-model techniques more available outside closed services: sparse expert routing, native image-and-text understanding, and extremely long advertised context. That gave teams more control over where inference runs and how a model is adapted.
It also exposed the real boundaries of local AI. Downloadable weights do not remove licensing conditions, infrastructure needs, evaluation work, or safety responsibilities. The release expanded the design space for private and customized systems, but successful deployment still depends on the surrounding software, hardware, data governance, and human oversight.