WizardLM-2 7B

WizardLM-2 7B is the smallest model in the WizardLM-2 family from the WizardLM team at Microsoft AI, announced on April 15, 2024 alongside WizardLM-2 8x22B and 70B. This entry links to lucyknada/microsoft_WizardLM-2-7B, a third-party re-upload of the original model card. The card describes it as a multilingual chat model built on Mistral-7B-v0.1, with 7B parameters (BF16) and the Apache 2.0 license.

Status and provenance

The re-upload's card does not say why the copy exists. When checked, the official microsoft/WizardLM-2-7B page on Hugging Face returned an authorization error (HTTP 401), so no first-party card could be read. Treat the weights as a third-party copy and verify files and provenance yourself. The card gives no deprecation notice or successor. The larger sibling in this catalog is WizardLM-2 8x22B.

Key Features

• Mistral-7B-v0.1 base: a small open model that runs on modest hardware.
• Vendor claims: the card says "WizardLM-2 7B is the fastest and achieves comparable performance with existing 10x larger opensource leading models."
• Human preference result: according to the card, WizardLM-2 7B is comparable with Qwen1.5-32B-Chat and surpasses Qwen1.5-14B-Chat and Starling-LM-7B-beta. The MT-Bench result is shown only as a chart image, so no scores are quoted here.
• Vicuna-style prompt format: multi-turn, starting with "A chat between a curious user and an artificial intelligence assistant," followed by USER: ... ASSISTANT: ...</s> turns.
• License: Apache 2.0.

The card does not state a context length.

Use Cases

• Small, self-hosted chat assistants and experiments.
• Comparing a 2024 7B instruct model against newer ones.
• Educational use of the Vicuna-style multi-turn template.

Limitation: the performance claims are the authors' own, and the model is a 2024 7B release.

Pricing

The weights are free under Apache 2.0. Cost is your own compute or a third-party host's price, which this page does not quote.

Getting Started

1. Read the card and confirm the license and files.
2. Load the weights with Transformers or a compatible server.
3. Format conversations with the Vicuna-style template from the card.

Frequently Asked Questions

Who made WizardLM-2 7B?

The WizardLM team at Microsoft AI, according to the card.

What is it based on?

Mistral-7B-v0.1.

Is the linked card official?

No. It is a third-party re-upload, and the official Microsoft page could not be read when checked.

What is its context length?

The card does not state one.

Alternatives

• WizardLM-2 8x22B: the larger WizardLM-2 sibling.
• Mistral 7B Instruct: Mistral's own instruct model of the same base family.
• OpenChat 3.5 7B: another Mistral-7B-based open chat model.

Conclusion

WizardLM-2 7B is a small Mistral-based chat model from Microsoft's WizardLM team, now reachable through third-party copies. Verify provenance before use, and compare it with newer small open models for new projects.