WizardLM-2 8x22B

WizardLM-2 8x22B is the most advanced model in the WizardLM-2 family, which the WizardLM team announced on April 15, 2024 on its release page together with a 70B and a 7B model. The page says the 8x22B and 7B models use Apache 2.0, while the 70B uses the Llama-2 Community License. The page does not name the base models. The official Microsoft repository on Hugging Face was not accessible when we checked, so this entry relies on the release page; the catalog website link points to a community mirror.

Key features

• Three sizes: 8x22B (most advanced), 70B and 7B (fastest), per the release page.
• Human preference: the team says the 8x22B model is slightly behind GPT-4-1106-preview and well ahead of Command R Plus and GPT4-0314 in its human-preference evaluation.
• MT-Bench: described as highly competitive with GPT-4-Turbo and Claude-3 on GPT-4-judged MT-Bench. These are the developers' own evaluations.
• Licenses: Apache 2.0 for 8x22B and 7B; Llama-2 Community License for 70B.
• Mirror caveat: community mirrors on Hugging Face exist, but we could not verify their contents from a first-party card.

Use cases

• Reading the original WizardLM-2 release claims when you meet them in older comparisons.
• Experiments with open instruction-tuned models from 2024 under Apache 2.0.
• Comparing 8x22B against the smaller 7B sibling.

Pricing and availability

The release page does not list prices. The 8x22B weights are described as Apache 2.0, but we could not open the official Hugging Face card on 2026-10-29, so confirm the license and file integrity yourself if you use a community mirror. Because we could not verify the model's current official availability, we do not label it live. The page reports no deprecation or successor.

Getting started

1. Read the WizardLM-2 release page for the original claims and licenses.
2. Locate the weights you intend to use and verify the license and source, since the official card was not reachable for us.
3. Run your own evaluation instead of relying on the published human-preference numbers.
4. Compare against current open models in this catalog before choosing.

FAQ

What is WizardLM-2 8x22B?

The most advanced model in the WizardLM-2 family, announced April 15, 2024 alongside 70B and 7B models.

What license does it use?

Apache 2.0 for the 8x22B and 7B models, per the release page.

How does it compare with GPT-4?

The WizardLM team reports it slightly behind GPT-4-1106-preview in human preference and highly competitive with GPT-4-Turbo on MT-Bench. These are the developers' own results.

Is there a smaller version?

Yes, see WizardLM-2 7B.

Alternatives

• WizardLM-2 7B: the smaller, faster sibling.
• Llama 3.1 70B Instruct: Meta's 70B open-weights instruct model.
• Mistral Large 4: a larger Mistral model.

Conclusion

WizardLM-2 8x22B is a 2024 open model whose claims come from its developers' release page, and whose official model card we could not reach. Treat the numbers as vendor claims, verify the weights you use, and see the release page and the models category.