WizardLM
The Evol-Instruct model family from Microsoft, inactive since 2024
Overview
WizardLM is an open instruction-tuned model family (WizardLM, WizardCoder, WizardMath) built by a Beijing-based Microsoft research group on Evol-Instruct, which evolves training instructions into harder variants. No new weights have shipped since early 2024; WizardLM-2 was withdrawn by Microsoft in April 2024, and in 2025 the group's lead researcher moved to Tencent Hunyuan.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with WizardLM
- The Evol-Instruct method and code for evolving training instructions into harder variants
- Ten downloadable WizardLM, WizardCoder and WizardMath models on Hugging Face
- Published papers on WizardLM, WizardCoder and WizardMath for instruction-tuning research
- Hosted WizardLM-2 8x22B through OpenRouter for comparison runs
- A documented case of a model release withdrawn over a skipped safety review
WizardLM's lasting contribution is Evol-Instruct, a widely reused idea for generating harder training instructions, and its 2023 models posted strong, self-reported scores for their size. The project has been inactive since 2024: WizardLM-2 was withdrawn within a day of release, and the lead researcher has since moved to Tencent.
Ten older models remain downloadable, with licences that vary from Llama 2 terms to none declared, and the Evol-Instruct data is non-commercial. Read it for the method; do not build on it.
What we found
WizardLM is a family of open instruction-tuned models built by a Beijing-based Microsoft research group around Evol-Instruct, a method published in April 2023 in which a language model rewrites training instructions into progressively harder versions, adding constraints, reasoning steps or combined tasks. The family grew to WizardLM for chat, WizardCoder for programming and WizardMath, fine-tuned from other companies' base models including Llama 2, Mistral and DeepSeek-Coder. Its own README reports strong results, such as 79.9 pass@1 on HumanEval for WizardCoder-33B-V1.1 and 83.2 on GSM8K for WizardMath-7B-V1.1; all of those are self-reported.
The project then stopped. Its last new weights appeared in January 2024. In April 2024 the team released WizardLM-2 and Microsoft pulled it within about a day, the team saying on X that a required toxicity-testing step had been missed; the repository records a deletion of the WizardLM-2 directory on 16 April. In May 2025 TechCrunch reported an X post by Can Xu, a senior researcher who led several of its projects, announcing that the WizardLM team had left Microsoft for Tencent's Hunyuan group, but 36Kr reported the post was deleted and Can Xu clarified that it was a personal move, not the whole team. Either way, nothing new has shipped under the WizardLM name.
What survives is downloadable. The WizardLMTeam account on Hugging Face still hosts ten models; the project's homepage is gone, and the GitHub repository has seen only README edits since mid-2024. WizardLM-2 survives as community re-uploads, and OpenRouter lists the 8x22B model at $0.62 per million tokens through a single third-party host.
Licences need checking model by model. The README badges declare code under Apache 2.0 and data under CC BY-NC 4.0, so the Evol-Instruct data is non-commercial, yet the repository carries no licence file. Several models inherit the Llama 2 licence and two declare none. Use WizardLM to study instruction evolution; use a maintained family to ship anything.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- WizardLM's Evol-Instruct creates instructions that are 4x more complex on average than the original training data, teaching models to handle requests that simple instruction datasets never include.
- The WizardLM family (WizardLM, WizardCoder, WizardMath) was produced by a team of 3 Microsoft Research interns in under 6 months.
- WizardLM-70B was the first open model to achieve GPT-3.5-level performance on the MT-Bench conversational evaluation.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including WizardLM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with WizardLM unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to WizardLM directly →
Spotted something out of date? Suggest an update →
Related reading
Health & Wellness
Fine-Particle Air Pollution Linked to Higher Hospital Admission Risk for Up to 21 Days, St...
Cybersecurity
Microsoft Patched a Record 1,411 Flaws in October and Suggested a New Way to Sort Them
Health & Wellness
A Guide for New Mothers Missed Its Main Goal but Nearly Doubled Primary Care Visits
Alternatives to WizardLM
More in LLMs & Chat
