[Qwen 3 32B Instruct](/models/alibaba--qwen-3-32b-instruct) and [Solar Pro 22B](/models/upstage--solar-pro-22b) are both mid-range instruction models, but they diverge in architecture: Qwen 3 32B is a dense transformer at 32B parameters; Solar Pro 22B uses Upstage's depth-up-scaling (DUS) architecture, which reportedly achieves competitive benchmark scores with fewer active parameters. Pricing reflects this — Solar Pro 22B typically comes in at $0.15–0.25/M tokens versus Qwen 3 32B's $0.20–0.35/M, a 20–30% discount. On reasoning-heavy benchmarks (MMLU, BBH), Qwen 3 32B holds a clear lead — roughly 5–8 points on MMLU versus Solar Pro 22B. The parameter count pays off on multi-turn conversations, complex instruction chains, and tasks requiring broad world knowledge. Qwen 3 32B also handles code generation and math better than most sub-30B models. Solar Pro 22B earns its place on Korean-language tasks — Upstage's training pipeline is notably strong in Korean NLP, and it outperforms comparably sized models on Korean benchmarks. For teams building Korean-language assistants or serving Korean-language users, Solar Pro 22B's quality-per-dollar ratio is hard to beat. It also works well for document processing workloads where DUS's efficiency advantage compounds across large batches. Pick Qwen 3 32B Instruct for general-purpose reasoning, English-dominant applications, and tasks where raw benchmark quality matters. Pick Solar Pro 22B if Korean-language performance is a requirement or if you want a 20–30% cost reduction on document-processing pipelines where the quality delta is acceptable.