Three of the most-discussed open-weights models in production, each representing a distinct tradeoff. DeepSeek V3.2 is a 671B-parameter MoE released May 2025 by a Chinese research lab, with approximately 37B active parameters per forward pass. That architecture keeps inference cost competitive with dense 70B models while delivering benchmark quality that rivals frontier proprietary offerings on code, math, and general reasoning. Context window is 128K. DeepSeek's own license applies, which is permissive for most commercial use but is not Apache, so enterprise legal teams will want to confirm before deployment. Llama 3.3 70B Instruct is Meta's 70B dense model released December 2024 with a 131K context window and the Llama 3 community license — the closest to Apache-permissive you will find at this quality tier from a major lab. It is the default recommendation for teams that want broad provider coverage, predictable licensing, and solid instruction-following without a licensing deep-dive. Benchmark improvements over 3.1 70B are genuine, not just marketing. Qwen 3 72B Instruct from Alibaba matches Llama 3.3 70B at the parameter tier with a competitive 131K context window and noticeably stronger multilingual performance across Chinese, Japanese, Korean, and Arabic. The Qwen commercial license covers production deployment. For global products serving non-English user bases, it often performs better on the workloads that actually matter. Pick DeepSeek V3.2 when you need the highest quality per inference dollar and can manage the licensing review. Pick Llama 3.3 70B for permissive licensing and the widest provider options. Pick Qwen 3 72B when multilingual breadth is a first-order product requirement.