WizardLM-2 (Microsoft)

tl;dr: WizardLM-2 is Microsoft’s next-gen chat family — the page ships two sizes (7B dense, 8x22B MoE) with 32K context. Both were last refreshed on the library 2 years ago, and the README’s “70B coming soon” never made it to a tag.

What it is

wizardlm2 is the umbrella tag Ollama ships for Microsoft’s WizardLM-2 release. The README on the library page describes it as “a next generation state-of-the-art large language model with improved performance on complex chat, multilingual, reasoning and agent use cases” and lists three variants:

• wizardlm2:7b — fastest; “comparable performance with 10x larger open-source models.”
• wizardlm2:8x22b — “the most advanced model, and the best opensource LLM in Microsoft’s internal evaluation on highly complex tasks.”
• wizardlm2:70b — listed in the README as “coming soon.” Not present in the library’s tag table as of this fetch.

The library table currently shows 22 model variants — mostly quantizations of the 7B and 8x22B bases.

Sizes and hardware

All variants are text-only. No vision, no audio tags on the page.

Pull

ollama run wizardlm2         # defaults to wizardlm2:7b
ollama run wizardlm2:7b
ollama run wizardlm2:8x22b

License

The ollama.com page does not state a license. Upstream WizardLM-2 is published under Microsoft’s own WizardLM-2 license terms — non-commercial research use by default, with commercial access negotiated separately. Check huggingface.co/microsoft/WizardLM-2 and the project’s blog post before any commercial deployment.

What the README actually says

Two short paragraphs and two reference links (blog post + HuggingFace collection). No benchmark numbers, no per-size quality discussion, no agentic evaluation details — even though the README claims “agent use cases” as a strength. The “Updated 2 years ago” badge matches the thin docs.

Gotchas

• The README promises a 70B that hasn’t shipped to the library. If you need a 70B-class Microsoft model today, look elsewhere.
• License is restrictive — Microsoft publishes WizardLM-2 under non-commercial terms. Confirm before any production use.
• Last refreshed on the Ollama library 2 years ago. Current Microsoft models live under phi4 and magistral tags on the same library, with active maintenance.
• 8x22B is an MoE — the active parameter count is ~22B, not the total ~141B. RAM/VRAM needs depend on which experts are loaded; check /library/wizardlm2/tags for the smallest variant that fits your box.
• 32K context is generous for the era, but no tool-calling tag is listed on the page; if you need native function calling, prefer a more recent model.

Related: Phi-4 for Microsoft’s actively-maintained small model, Qwen3 for a current multilingual default, Ollama for the runtime.

Untested here — facts from ollama.com/library/wizardlm2, the WizardLM-2 blog post, and the Microsoft HuggingFace collection fetched 2026-09-10.