Meta's Llama 4 model family has delivered the multimodal capability improvement that enterprise AI teams were waiting for, but the deployment reality is more complicated than the benchmark announcement suggested. Llama 4 Scout (17B active parameters, 109B total with mixture-of-experts routing) and Llama 4 Maverick (17B active, 400B total) both show 35-45% improvements over Llama 3 on image understanding, document analysis, and mixed text-image reasoning benchmarks.
The benchmark that attracted the most attention: Llama 4 Maverick scores 73.5% on MMMU (Massive Multitask Multimodal Understanding), placing it ahead of GPT-4o (69.1%) and Claude 3.5 Sonnet (68.0%) on this specific evaluation. Meta presented this as evidence that open-source models have reached frontier multimodal capability — a claim that is accurate on this particular benchmark, with important caveats about benchmark saturation and task coverage.
Enterprise adoption is running into a practical constraint. Llama 4 Maverick's 400B total parameter count requires approximately 600GB of VRAM at FP16 precision — roughly eight A100 80GB GPUs or five H100 80GB GPUs just for the model weights. At current spot market prices, that infrastructure runs $3-$5 per hour, which is comparable to commercial API pricing for models of similar capability. The "free" in open-source requires significant infrastructure investment to realize.
Quantized versions (INT4 Llama 4 Maverick) reduce the hardware requirement by roughly 70%, enabling deployment on 2-3 H100s, but at a performance cost of approximately 8% on benchmark tasks. For most production use cases, quantized Maverick represents a reasonable trade-off.
Llama 4 Scout, the smaller variant, runs on a single A100 80GB GPU in FP16 — this is the version seeing the most enterprise evaluation activity for private deployment scenarios where data privacy justifies the infrastructure cost.
Meta's release velocity (Llama 4 following Llama 3 in under 12 months) has made it the reference architecture for teams exploring open-source alternatives to commercial APIs.