Llama 3.1 and Llama 4 represent different architectures and release vintages. Meta released Llama 3.1 on 23 July 2024, expanding context to 128K and adding a 405-billion-parameter variant. The first Llama 4 Scout and Maverick models were announced on 5 April 2025 as native multimodal mixture-of-experts systems. They are often casually called “open source,” but selecting weights for an enterprise workload requires reading their specific community licenses and acceptable-use terms.
Hardware, not just parameter count
Llama 3.1 includes dense models in different sizes. Llama 4 uses active versus total parameters, so a 17B-active model is not equivalent to a dense 17B checkpoint in memory needs. Quantization and tensor parallelism can change throughput and precision. Meta states that Llama 4 Scout may fit on a single H100 under an Int4 configuration; that is a deployment-specific vendor claim, not an endorsement of every local runner.
Multimodal inputs
Llama 4 introduced image-and-text capabilities; that does not confer image generation or native streaming speech output. Validate each checkpoint and tokenizer: “Scout,” “Maverick,” and a future “Behemoth” preview must have separate records. Meta’s touted context windows should be treated as maximum supported input length under particular serving conditions, not guaranteed retrieval at all positions.
Licensing and sovereignty
Open weights let teams operate their own serving stack and control storage location, but they do not automatically provide unlimited redistribution, trademark use or training-data rights. For a regulated deployment, document license text and revision, model hash, hosting provider, storage and log retention. Model downloads can create supply-chain obligations; use signed artifacts and reproducible packaging when available.
Decision framework
Choose dense Llama 3.1 where mature tooling and a stable 2024 checkpoint matter. Investigate Llama 4 for text-plus-image workloads with supported hardware. Benchmark localized language and robustness on held-out data. Distinguish vendor-reported benchmark wins from results independently reproduced in your environment.
What to verify before adopting this model or method
Use the exact release ID and API endpoint; confirm your geographic processing location, entitlement tier, model lifecycle, feature flags, context and output limits from the provider console. Run a reproducible evaluation on your own data and retain human oversight for consequential actions. Where a specification, price, commercial license or entitlement is not verified for your deployment, this article deliberately leaves it unasserted instead of filling a plausible number.
Documentation clarification (25 October 2026)
Meta’s April 5, 2025 release announced Llama 4 Scout and Maverick as downloadable open-weight models while describing Behemoth as a preview not released at launch. “Open weights” does not remove license conditions or guarantee a one-GPU deployment at arbitrary precision.
Sources & Further Reading
- ai.meta.com — documentation/reference (supplier-provided reference; not independently checked in this pass)
- Meta Llama 4 launch (retrieved from first-party public documentation on 2026-10-25; this is not a deployment test)
Source documentation establishes vendor-published facts, not independently measured model quality, tenant access, contractual rights or current prices. Use a dated pricing ledger and an account-specific entitlement check for deployment decisions.