WizardLM-2 8x22B

WizardLM-2 8x22B sits at the top of the WizardLM-2 family introduced in mid-April 2024, a three-tier lineup that also includes a 70B reasoning-focused sibling and a 7B speed-oriented variant. The WizardLM team framed the release as a next-generation open-source large language model family with stated improvements across complex chat, multilingual understanding, reasoning, and agent-style tasks, and they positioned the 8x22B mixture-of-experts configuration as their most capable entry for highly complex workloads. Weights for the family are hosted on the microsoft/ namespace on Hugging Face, signaling a collaboration between Microsoft and the WizardLM research group rather than a purely independent open-source effort, and the project ships with a public release blog that documents the family’s design goals and evaluation setup.

In the WizardLM team’s own internal evaluation, WizardLM-2 8x22B is described as the best open-source LLM they benchmarked on highly complex tasks and as showing competitive performance against leading proprietary systems on the MT-Bench human-preference suite, alongside strong results from the smaller 70B and 7B siblings. The announcement emphasizes that the family was designed to close the gap with closed frontier models on reasoning and instruction-following while remaining fully open for self-hosted deployment and fine-tuning. For practitioners, the practical fit is large-scale open-weight experimentation: organizations that want to run a top-tier open model for complex chat, multilingual, and agent workflows without paying per-token API fees can build on the 8x22B mixture-of-experts weights and tailor them to domain-specific tasks.