Two mixture-of-experts models and one dense model, all competitive at the upper end of the open-weights market. DeepSeek V3.2 is the newest and most capable — a 671B-parameter MoE from May 2025, routing approximately 37B active parameters per token, with benchmark performance matching or exceeding frontier proprietary models on code and math. The context window is 128K. DeepSeek's own license applies; commercial teams should review it before production deployment. At this quality tier, it has effectively set a new cost-adjusted baseline that makes older large MoE models harder to justify on quality grounds. Mixtral 8x22B Instruct is Mistral's April 2024 release, 141B total parameters with roughly 39B active per forward pass and a 64K context window. Apache 2.0 license makes it the cleanest option in this comparison for self-hosted commercial deployment without legal overhead. Quality on English reasoning and coding tasks remains competitive for its release year, and the base architecture is stable enough that fine-tuned variants (including WizardLM-2 8x22B) have extended its utility. Its main disadvantage is age: two newer generations of open models have raised the quality bar. Qwen 3 72B Instruct is Alibaba's April 2025 72B dense model with a 131K context window and Qwen commercial licensing. It punches above its parameter count on multilingual evaluation suites and matches or beats Mixtral 8x22B on English benchmarks at lower active-parameter cost. For teams that process mixed-language workloads or specifically serve CJK markets, it offers a clear advantage. Pick DeepSeek V3.2 for maximum cost-adjusted quality with licensing review. Pick Mixtral 8x22B when Apache permissiveness and stable, audited architecture are the primary requirements. Pick Qwen 3 72B for multilingual production workloads where dense 72B quality suffices.