Three MoE instruction models from 2024, all routing 36–42B active parameters per token, each representing a different publisher's approach to large open-weights inference. DBRX Instruct from Databricks launched March 2024 with 132B total parameters and 16 fine-grained experts, beating Mixtral 8x7B on most benchmarks at the time. The 32K context window is its main limitation relative to the other two here. The Databricks Open Model License is permissive for most commercial use but not OSI-approved — worth a legal review before deployment. Its clearest remaining use case is teams running in the Databricks ecosystem who want a native Databricks model for integration reasons. Mixtral 8x22B Instruct is Mistral's 141B-total-parameter release from April 2024, with roughly 39B active per pass and a 64K context window. It was a significant step up from Mixtral 8x7B on reasoning and multilingual tasks, and it remains the base architecture that WizardLM-2 8x22B builds on. The Apache 2.0 license is the cleanest of the three for commercial deployment and redistribution. WizardLM-2 8x22B is Microsoft Research's April 2024 fine-tune of the Mixtral 8x22B base using Evol-Instruct, which produced measurably stronger multi-turn conversational benchmark scores at release. The WizardLM 2 Community License carries non-standard attribution clauses that require review before commercial deployment. At 64K context and the same active parameter count, it slots in above the base Mixtral 8x22B on conversational quality. Pick DBRX when you are already in the Databricks stack. Pick Mixtral 8x22B when Apache licensing and self-hosting flexibility matter. Pick WizardLM-2 8x22B for multi-turn conversational tasks where you can accommodate the attribution requirements.