Three compact models from labs outside the top-tier American AI ecosystem, each reflecting a different commercial strategy. Mistral Small 3 is Mistral AI's January 2025 small-model release, sitting in the sub-25B range with a 32K context window and Mistral's Apache 2.0 license — notably more permissive than Mistral Large 2's Research License. It targets cost-sensitive production tasks where Mistral's instruction-tuning quality matters but the 123B scale is unnecessary. Function calling and JSON mode carry over from the larger models. For European teams that want Mistral's multilingual quality at reduced inference cost, this is the practical entry point. Solar Pro 22B is Upstage's October 2024 release, a 22B dense model designed explicitly for single-GPU deployment fitting on one A100 or H100 under int8 quantization. That engineering choice reflects a specific customer profile: teams that process mixed Korean-English enterprise documents without multi-GPU overhead. The CC-BY-4.0 license is unusually permissive at this scale — attribution only, no share-alike, no commercial restrictions. The 4K context window is a real deployment constraint, and Upstage's provider coverage remains limited compared to Mistral or Meta alternatives. Yi 1.5 34B Chat from 01.AI is a May 2024 34B bilingual model under Apache 2.0, covering English and Chinese with benchmark scores that held up reasonably at release. The 4K context ceiling matches Solar Pro's constraint and rules out both models from most RAG architectures without tight chunking strategies. As of mid-2026, Qwen 2.5 and DeepSeek V2 have moved past Yi 1.5 34B on bilingual Chinese quality, but Apache licensing with no commercial restrictions keeps it relevant for specific use cases. Pick Mistral Small 3 for European multilingual workloads where Apache 2.0 freedom and Mistral's instruction quality are the priorities. Pick Solar Pro 22B for Korean-English bilingual enterprise document processing on a single GPU. Pick Yi 1.5 34B Chat where Apache-licensed Chinese-English bilingual support is required and 4K context is acceptable.