Meta switched from a dense architecture to mixture-of-experts with Llama 4, and that technical change has direct implications for what hardware you need and what license you're actually agreeing to.
Unlike Llama 2 and 3, which used dense transformer architectures (all parameters activate on every forward pass), Llama 4 adopted Mixture of Experts (MoE): the model has a large number of total parameters, but only a subset ("experts" activated by a router) gets used per processed token. This is the same family of techniques used by Mistral's Mixtral and (speculatively) GPT-4.
The practical advantage is you can have a model with hundreds of billions of total parameters but an inference cost comparable to a much smaller dense model, because only the active parameters consume compute per token. The downside is the full model still needs to be loaded in memory (VRAM or RAM) even though not all of it gets used at each step, so the savings are in compute, not total memory footprint.
Meta released Llama 4 in two main variants aimed at different target hardware. Scout is the lighter-in-active-parameters variant, designed to run on a single high-end GPU (think H100), with an aggressively extended context window -- Meta positioned it with support for up to 10 million tokens of context, though in practice effective performance (real recall over long context) degrades well before reaching that theoretical limit, a pattern repeated across almost every model that announces "extreme" contexts.
Maverick is the higher-capacity variant, with more total experts and better performance on reasoning and code benchmarks, but requires multi-GPU to run in FP16/BF16 -- it's not viable on a single consumer node, and its realistic deployment requires tensor parallelism across multiple GPUs with NVLink or equivalent interconnect.
Llama doesn't use an OSI-approved license like MIT or Apache 2.0 -- it uses the "Llama Community License," which imposes specific restrictions: companies with more than 700 million monthly active users need a separate commercial license from Meta, and there are mandatory attribution clauses (you must include "Built with Llama" in derivative products). This makes it "open-weight" (downloadable, modifiable weights) but not "open-source" in the strict Open Source Initiative sense -- a distinction that does matter legally for companies planning to commercially redistribute or serve the model at large scale.
For the vast majority of Latin American companies that would use Llama 4 internally or in their own product without hitting that user threshold, the license is permissive and free -- the real point of attention is only fine-tuning and redistributing derivative weights, where the attribution clauses do apply.
Compared to other open-weight models of similar weight class (Alibaba's Qwen2.5/Qwen3, Mistral Large, DeepSeek-V3), Llama 4 doesn't dominate overwhelmingly on every benchmark -- the narrative of "Meta always leads open-weight" weakened considerably during 2024-2025 with the strong entry of Chinese models (Qwen and DeepSeek), which on several code and mathematical reasoning benchmarks beat comparably-sized Llama variants. The real choice today depends more on the specific use case and the tooling ecosystem (fine-tuning, quantization, library support) than on a single benchmark.
Given the size of Llama 4's variants, running it locally on a single consumer GPU is only viable with aggressive quantization and probably only for Scout. For most enterprise cases, the practical path is: Ollama or vLLM for Scout quantized in GGUF on a 24-48GB GPU, or managed inference providers (Groq, Together AI, Fireworks) for Maverick, where the cost of your own infrastructure doesn't justify itself against the service's per-token price.
Llama 4 makes sense when you already have Meta-adjacent infrastructure (using PyTorch, familiarity with fine-tuning tooling like torchtune), when you need genuinely extended context for very long documents, or when the long-term support guarantee of an actor like Meta weighs into your risk decision. For most RAG or agent projects with normal-sized documents, smaller, more efficient models (Qwen2.5 14B-32B, Mistral Small) usually give a better cost-benefit ratio than deploying the full Llama 4 family.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel