WizardLM-2 8x22B in 2026: A Pulled Release That Never Went Away

Most models in an API catalogue have a clean origin: a lab published weights, a provider loaded them. WizardLM-2 8x22B does not. Microsoft's WizardLM team released it on 15 April 2024, withdrew it from Hugging Face and GitHub within hours, and there is no sign that the official 8x22B repository ever came back. Copies had already spread, and those copies are what providers have served ever since.

That history matters more than any benchmark when you decide whether to put it in a product today. This guide covers what the model is, why it was popular, what the withdrawal means in practice, and how it compares with the options available now.

What the model actually is

WizardLM-2 8x22B is a fine-tune of Mixtral 8x22B, Mistral's mixture-of-experts base model, released within days of that base becoming available. The surviving model card lists 141B parameters in a mixture-of-experts layout, the base as Mixtral-8x22B-v0.1, and the licence as Apache 2.0. It was the flagship of a family of three: an 8x22B, a 70B built on Llama 2 (and so under the Llama 2 licence), and a 7B.

If you want the details of the base architecture and how expert routing affects serving cost, the Mixtral 8x22B Instruct guide covers Mistral's own instruction-tuned version of the same base. That is the most natural comparison, because both start from identical weights.

The prompt format is Vicuna's, not Mistral's: alternating USER: and ASSISTANT: turns, with each assistant turn closed by an end-of-sequence token. Through an OpenAI-compatible API the provider applies this for you, but if you self-host, check the template: Mixtral's chat template is not the format this fine-tune was trained on.

Why people liked it

The interest was in the training method rather than the base. The WizardLM team described a fully synthetic pipeline: reweighting source data by attribute distribution, an updated Evol-Instruct that uses multiple agents to generate harder instructions, an Evol-Answer step that rewrites responses repeatedly, models critiquing each other's output, then supervised learning, staged DPO and a reinforcement-learning stage with separate instruction-quality and process-supervision reward models.

The claims were bold for an open model in April 2024. The team reported that in human preference comparisons the 8x22B was slightly behind GPT-4-1106-preview and well ahead of Command R Plus and the original GPT-4-0314. On MT-Bench it was described as competitive with GPT-4-Turbo and Claude 3, though the scores were published only as a chart. For conversational writing and long, well-structured answers, it earned a following that outlasted the takedown.

More neutral measurements are less flattering. The Hugging Face Open LLM Leaderboard shows an average of 32.61 on its normalised scale, with 22.28 on MATH Level 5 and 17.56 on GPQA. Those are reasonable for 2024 and weak by the standard of current open models.

The withdrawal, and why it still matters

The stated reason for pulling the release was procedural: the team said it had skipped the toxicity testing Microsoft required before publishing, and intended to run it and re-release. There is no sign that an official re-release of the 8x22B followed. OctoAI later published third-party toxicity testing of the model, but that is not the same as the vendor completing its own review.

Three practical consequences follow:

1. Provenance. Every copy in circulation is a re-upload, and there is no official repository left to verify a checksum against. For a regulated environment, that alone may rule it out. Security review for open-weight models covers what a reasonable check looks like.
2. Safety tuning. The model never received the safety pass its own authors said it needed. Put a classifier in front of it if users can reach it directly.
3. No updates. There will be no fixes, no new version and no vendor support. What you see is what it will always be.

What it costs, and what it competes with now

At the time of writing OpenRouter lists WizardLM-2 8x22B at $0.62 per million tokens for both input and output, with a context window of roughly 64K tokens and JSON mode listed as supported. A symmetric price is unusual; most current models charge several times more for output than input, so for output-heavy work it compares better than the headline suggests.

The honest comparison, though, is capability per pound. Current open models with far stronger reasoning and coding scores are priced in the same range per million tokens, with much longer context and active maintenance. Cost-adjusted model scoring lays out how to rank models on value rather than raw price.

Should you still use it

There are two legitimate reasons. The first is continuity: an existing product whose prompts, tone and evaluation set were tuned on WizardLM-2, where switching costs real work. The second is style: some teams prefer its long-form conversational voice for creative or role-play work and have measured that preference.

For new work, the case is weak. If you are evaluating it anyway, run your own task set against Mixtral 8x22B Instruct and one current open model, score blind, and include a few safety-sensitive prompts. Building a 20-task eval covers the method. If WizardLM-2 does not clearly win on your own work, the provenance and maintenance questions settle it.