Three large open-weights models with strikingly different intended roles. Llama 3.1 405B Instruct is Meta's July 2024 maximum-scale release — 405 billion dense parameters, 131K context, and the Llama 3 community license. It is positioned as the quality ceiling for openly licensed dense models and serves as a teacher model for distillation, a general-purpose top-tier assistant, and a capability reference for researchers. Hosting requires significant multi-GPU infrastructure; Lambda Labs and a handful of other providers carry it. The license is the most permissive of the three here by a notable margin. Nemotron-4 340B Instruct is NVIDIA's June 2024 flagship, a 340B dense model with a 4K context window concentrated on NVIDIA's NIM service. Its design purpose is synthetic data generation — NVIDIA built it as a high-quality teacher model for producing instruction fine-tuning datasets, not as a chat backend. The 4K context ceiling is severe for most other applications. The NVIDIA Open Model License is not OSI-approved, and commercial use outside of synthetic data generation is an awkward fit. If your pipeline involves generating diverse synthetic instruction data at scale, this is the obvious specialist. For anything else, the alternatives are more practical. WizardLM-2 8x22B from Microsoft Research is a 141B MoE (39B active) with a 64K context window, released April 2024 as an Evol-Instruct fine-tune of Mixtral 8x22B. It excels on multi-turn conversational evaluations, which is where the Evol-Instruct methodology concentrates quality. The WizardLM 2 Community License is permissive in practice but carries non-standard attribution terms. Pick Llama 3.1 405B for maximum open-weights capability at frontier-scale with clean licensing. Pick Nemotron-4 340B exclusively for synthetic data generation pipelines where NVIDIA NIM access is available. Pick WizardLM-2 8x22B for cost-efficient MoE conversational quality where Mixtral-class infrastructure is already in place.