Tencent hires WizardLM team, a Microsoft AI group with an odd history

WizardLM, a Beijing-based Microsoft AI probe group, appears to person joined Tencent, the Chinese institution that owns WeChat and blockbuster games similar PUBG Mobile.

In a post connected X connected Tuesday, Can Xu, a elder AI researcher who led a fig of projects wrong WizardLM, said that helium and the squad near Microsoft to articulation Hunyuan, 1 of Tencent’s AI improvement orgs. Over the past fewer months, Hunyuan has released AI models including video and 3D entity generators.

WizardLM has already released a Hunyuan-branded model, successful fact: Hunyuan-TurboS 0416. In an X post, Qingfeng Sun, who claims to beryllium a “co-founder” of WizardLM, says that Hunyuan-TurboS 0416 outperforms “open” AI models similar Google’s Gemma 3 series.

It’s not wide however galore researchers are inactive a portion of WizardLM oregon erstwhile the squad might’ve near Microsoft. We’ve reached retired to some Tencent and Microsoft for comment.

WizardLM, formed respective years ago, has a spot of a unusual history.

Last April, the team, past nether Microsoft, released a household of AI models, WizardLM-2, that it said was competitory with OpenAI’s GPT-4. But aft lone a day, Microsoft pulled WizardLM-2 from the web, saying that the models hadn’t undergone “toxicity testing.”

“We accidentally missed an point required successful the exemplary merchandise process — toxicity testing,” the WizardLM squad wrote successful a post connected X connected April 16, 2024. “We are presently completing this trial rapidly and past volition re-release our exemplary arsenic soon arsenic possible.”

The deletion came a small excessively late. Users rapidly reuploaded the archetypal WizardLM-2 models arsenic good arsenic customized, fine-tuned versions.

Meanwhile, successful a station connected the AI dev level Hugging Face, Hugging Face CEO Clément Delangue said that Microsoft’s move, which besides wiped retired different WizardLM models, harmed the Hugging Face assemblage by breaking a fig of unfastened root projects.

“[The WizardLM] models were collectively downloaded implicit a 100 1000 times a month,” Delangue wrote astatine the time. “We apologize for the inconvenience & are trying to get successful interaction with the writer [and] Microsoft successful bid to effort to find a bully solution for assemblage members.”

At Tencent, it seems WizardLM volition support doing what it was doing, much oregon less: processing and releasing AI models. Tencent recently restructured its Hunyuan team, establishing 2 caller units and upping its AI infrastructure spending.

Tencent attributed its 8% year-over-year maturation successful Q1 2025 to the company’s AI investments. Tencent has said that it plans to enactment 90 cardinal yuan, oregon astir $12.49 billion, toward superior expenditures this year, overmuch of which volition substance its AI efforts.