WizardLM 2 8x22B

WizardLM 2 8x22B is a Mixture-of-Experts model developed through Microsoft's WizardLM research program, built on the Mixtral 8x22B architecture and fine-tuned using the Evol-Instruct methodology. This training approach progressively rewrites instructions into more complex variants, producing a model that handles nuanced, multi-step directions with high fidelity.

The model targets complex instruction following, long-form reasoning, and structured output generation. Its MoE backbone activates a subset of experts per token, giving it strong capacity without proportionally high inference costs. It performs well on coding assistance, writing tasks, and analytical queries requiring detailed responses.