Scale versus specialization. [Qwen 3 72B Instruct](/models/alibaba--qwen-3-72b-instruct) is Alibaba's latest 72B general-purpose model — competitive on coding and multilingual benchmarks, with a 128K context window and pricing typically in the $0.20–$0.50/1M range. DeepSeek R1 is a reasoning-specialized model trained with reinforcement learning for chain-of-thought derivation, running $0.50–$1.50/1M with additional thinking-token costs on complex tasks. The pricing gap is 2–4x, which means Qwen 3 72B Instruct wins by default on any workload where both models produce acceptable accuracy. That's a meaningful share of production use cases — RAG pipelines, classification, summarization, and tool-augmented agentic tasks where the 72B parameter base is sufficient. [DeepSeek R1](/models/deepseek--deepseek-r1) pulls ahead on tasks that explicitly require extended reasoning: AIME-level math, multi-step code debugging where you need the model to articulate why an approach fails, or scientific Q&A where intermediate derivation steps affect downstream tool calls. On MATH-500 and similar benchmarks, R1's scores are significantly higher than any 72B model regardless of training approach. Qwen 3 72B Instruct earns its place for multilingual enterprise workloads — Chinese-English bilingual processing, mixed-language document pipelines, or any application where Alibaba's multilingual pretraining corpus shows. It's also a credible choice for high-throughput coding assistance at lower cost than R1. Pick DeepSeek R1 if the task requires verifiable multi-step reasoning and cost is secondary. Pick Qwen 3 72B Instruct for multilingual workloads, cost-sensitive inference, or general-purpose tasks where 72B capability is sufficient.