Llama 3.1 405B, 70B, and 8B Instruct are the three primary tiers of Meta's July 2024 Llama 3.1 release, all under the Llama 3 community license and all carrying a 131K context window. The 3.1 generation was notable for shipping long context at the 70B tier for the first time in the Meta lineup. Together, these three models represent the canonical open-weights cost-quality ladder from light inference to frontier capability. The 8B is the volume workhorse. MMLU in the low-to-mid 70s, solid general-purpose instruction following, sub-$0.20 per million tokens on most providers. It handles short-form generation, summarization, classification, and lightweight coding tasks with reasonable reliability. Multi-step reasoning and complex agent tasks expose its limits quickly. The 70B is where quality becomes production-grade for most use cases. MMLU around 79-80, meaningfully better instruction-following and multi-turn coherence than the 8B. The 131K context makes it the first point in this family where long-document analysis is tractable without chunking. For new deployments at this tier, Llama 3.3 70B released December 2024 is generally a better choice, having improved instruction tuning at the same hardware footprint. Running 3.1 70B in 2026 is usually a pinned-checkpoint decision. The 405B sits at Meta's open-weights capability ceiling. It handles complex reasoning chains, extended code generation, and long-form analysis that visibly saturates 70B models. Serving 405 billion parameters requires multi-GPU infrastructure; provider availability is thinner than at the 70B and 8B tiers, and per-token pricing is substantially higher. It earns its cost for genuinely hard tasks, but is not the right default for production volume inference. Pick the 8B for high-throughput, cost-sensitive workloads. Pick Llama 3.3 70B for most new production deployments at the mid-tier. Use 405B when task complexity visibly saturates the 70B and the infrastructure overhead is justified.