WizardLM 2

WizardLM 2 was released by Microsoft in 2024 but was subsequently withdrawn from public distribution amid concerns about data licensing. The model weights were briefly available on Hugging Face, but access has been restricted, so users should check current availability rather than assuming open access.

Microsoft reported that the WizardLM 2 8x22B variant achieved performance competitive with leading proprietary models on several benchmarks. However, as an open-weights mixture-of-experts model, it has a smaller effective parameter footprint per token than GPT-4, and independent evaluations have shown mixed results depending on the task.

WizardLM 2 was removed from public release shortly after its debut in 2024 due to concerns that the training data may have included content from other models in ways that raised licensing or attribution issues. Microsoft has not provided a detailed public explanation, and the model's long-term availability remains uncertain.

WizardLM 2 is the second-generation series released in 2024, featuring improved instruction-following capabilities and a mixture-of-experts architecture in its larger variants. The original WizardLM, released by Microsoft Research, used Evol-Instruct methodology to generate training data, and the second generation builds on that approach with updated base models and training techniques.