Xiaomi Just Put Its AI Chief at the Top of the Company. Here’s What That Signals.

1. Two announcements, one day

On September 22, 2026, Xiaomi sent an internal promotion list. Luo Fuli, 31, head of the company’s large-model team, moved to Level 22. The same day, Xiaomi released and open-sourced the MiMo-V2.6 series.

A promotion notice is organizational language. A model release is product language. Saying both on the same day ties the two together: the company is putting its highest job level behind the work.

People close to Xiaomi say Level 22 is the top of the job-grade system. Further moves are about role, not number. Xiaomi built the system in 2019. Caijing reported at the time that it ran from Level 13 to Level 22: specialists around 13, managers 16 to 17, directors 19 to 20, vice presidents 22. Lei Jun sits outside it.

Luo Fuli confirmed she joined Xiaomi on November 12, 2025. On WeChat Moments, she wrote: “Intelligence will eventually move from language to the physical world. I am at Xiaomi MiMo, with a group of creative, talented, and sincerely passionate researchers, committed to building such a future and racing toward the AGI we envision.” Ten months later, Level 22 made that positioning official.

2. The July restructuring: who got the brain

If you read the promotion only through Luo Fuli’s resume, it becomes a story about a young star. That story explains little. The better question is why Xiaomi gave its highest level to the head of a large-model team at this moment.

In July 2026, Xiaomi reorganized the technical system behind Xiao Ai. It split into three parts. Foundation models went to Luo Fuli’s MiMo team. On-device work went to the OS teams for phones and cars. Cloud engineering went to Luan Jian’s team. The foundation model is the brain. The on-device layer is the hands and feet. The cloud is the pipes. Luo Fuli got the most upstream piece.

Luan Jian had led Xiaomi’s AI Lab large-model team since April 2023. After Luo Fuli joined at the end of 2025 and took over that business, he shifted to Agentic AI algorithms. After the July adjustment, his team became the “Harness” in the AI chain: it turns the large model into a stable interface so product teams can call it. 36Kr put it this way: Luan Jian’s team went from horse tamers to engineers fitting saddles and reins. The MiMo foundation built by Luo Fuli’s team is the horse.

Wang Gang, who had led the Xiao Ai team, moved to robotics. His nearly decade-long run with Xiao Ai ended.

People close to the business told 36Kr that the direct trigger was the Doubao phone’s popularity last year. After that, Xiaomi began rebuilding its terminal R&D around AI. In the old setup, Xiaomi devices relied on standardized cloud capabilities built by the Xiao Ai team. But on-device sensors do real-time perception and handle large amounts of data. Each product has different chips and compute. The cloud team could not process that. The old model could not carry the AI promise of the Human-Car-Home strategy.

After the adjustment, MiMo’s internal position changed. A person familiar with the matter told Leiphone that MiMo must enter phones, cars, home devices, wearables, and robotics as a long-term strategy. It has to become an underlying capability every business line can call. “If every line builds its own AI capabilities, it will easily end up with one assistant in the car, another on the phone, and yet another logic for home devices. Once the experience is fragmented, the ‘Human-Car-Home’ ecosystem cannot even get off the ground.”

The restructuring was meant to solve that. Unifying foundation-model work moved Xiaomi from “every line does its own thing” to “one foundation plus scenario adaptation.” Luo Fuli’s Level 22 is the organization confirming that direction.

3. MiMo-V2.6: the basis for promotion

A job-level promotion needs output. For Luo Fuli, that output is MiMo-V2.6.

The series launched on September 22 with two native omni-modal models. Pro handles complex coding, agentic tasks, and professional work. Flash balances performance, efficiency, and cost. Xiaomi also launched MiMo-V2.6-Pro-UltraSpeed, which can output up to 20 times faster than Pro.

On the Artificial Analysis Intelligence Index v4.3.2, MiMo-V2.6-Pro scored 46. Kimi K3 scored 44. GLM-5.3 scored 45. That made MiMo-V2.6-Pro the highest-ranked open-weight model on the AA index at the time. Its predecessor, MiMo-V2.5-Pro, scored 26, a 20-point jump in one generation. On Toolathlon-verified, Pro scored 76.9, above GPT 5.6 Sol’s 74.9. On JobBench, it scored 62.0, behind only Claude Opus 5’s 65.7. It still trails GPT-6 Astra and Claude Fable 5.1, both at 53 on the AA leaderboard. Its 34.9 on Terminal Bench 4.0 is well behind leading closed-source models.

For an open-source model, 46 plus the price may matter more than individual rankings. MiMo-V2.6 kept V2.5’s API pricing. Pro costs 3 yuan per million input tokens and 6 yuan per million output tokens. Flash costs 1 yuan and 2 yuan. At similar intelligence levels, that is a fraction of leading overseas models.

Xiaomi also made the training process public. Luo Fuli livestreamed MiMo-V2.6’s reinforcement learning training on X. She showed token consumption, training costs, and cumulative expenses in real time. The two versions together cost about $31,000 per hour, more than 200,000 yuan. Pro’s training cost about $2.62 million. Flash’s cost about $850,000. Training took less than six days.

Luo Fuli wrote on X that the team was studying how far reinforcement learning could go. The question was whether large-scale RL gains would generalize beyond the training distribution. On DeepSWE v1.1, a long-horizon software engineering benchmark not used in training, Flash rose from 48.8 to 65.68. Pro rose from 58.4 to 72.57. Xiaomi concluded that large-scale RL still has relatively high sample efficiency and that the model learned some general long-horizon execution.

The spending produced a model with measurable gains. That gave the promotion its basis.

4. 60 billion and 16 billion: AI as foundation

In March 2026, Lei Jun said at the spring launch that Xiaomi would invest more than 60 billion yuan in AI over three years. AI R&D and capital spending this year would reach 16 billion yuan. First-quarter R&D spending was 9 billion yuan, up 33.4% year-on-year. R&D staff totaled 26,048. Full-year R&D spending is expected to exceed 40 billion yuan.

On the earnings call, Lu Weibing said the core strategy remains strengthening foundation-model capabilities and using self-developed foundation models to power Xiaomi’s Human-Car-Home strategy. He also said Xiaomi will not rush to monetize AI tokens and will move with a pragmatic attitude.

Xiaomi’s AI strategy has a structural difference from other vendors: it has a terminal matrix across Human-Car-Home. Phones, cars, home appliances, and other devices are entry points for AI to enter the physical world. Xiaomi HyperOS rebuilt the system layer and connected people, cars, and homes. HyperOS 4 carries the MiMo large model and the fully AI-enabled “Super Xiao Ai 2.0.” It enables cross-platform, cross-scenario, and cross-device collaboration.

MiMo’s deployment scenarios are built in. Xiaomi does not need outside customers to validate model capabilities the way pure model companies do. Its phones, cars, and IoT devices are testing grounds and deployment environments. But that also means MiMo’s ceiling sets the ceiling for AI experience across Xiaomi’s entire device network. If the foundation model is weak, every scenario above it suffers.

Luo Fuli’s team is small. Lei Jun said the MiMo model was developed by Xiaomi’s core technical team, Core Team. The average age is 25. Tsinghua and Peking University graduates make up more than 60%. PhDs make up 55%. The youngest researcher is a 19-year-old intern. About 100 people work across the full chain: data collection, quality review, infrastructure, pre-training, post-training, algorithm R&D, and product deployment.

Supporting a large device network with a small team puts heavy demands on the lead. Level 22 recognizes Luo Fuli’s ability and confirms the team’s strategic position inside Xiaomi.

5. The talent war behind the promotion

The promotion also lands in a fierce talent market.

Maimai reported that average monthly pay for AI roles reached 60,738 yuan. AI scientists and leads averaged 137,153 yuan per month. Algorithm researchers and large-model algorithm roles generally earned around 70,000 yuan per month. Senior large-model algorithm experts earn 1 million to 2 million yuan a year. Multimodal algorithm engineers earn 600,000 to 1.5 million yuan. Embodied intelligence algorithm engineers can earn up to 2 million yuan.

Supply is tighter than pay suggests. The hardest roles to fill are core R&D talent: foundation model algorithm experts, pre-training experts, multimodal architects, AI Infra experts for very large models, and AI agent architects. These roles have few candidates, and top people move rarely. A person doing pre-training at a major internet company said the scarce thing is taste for data. Large-model capability has four pillars: architecture, data, infra, and evaluation. Data currently has the greatest impact on model performance.

Poaching has become mutual. Xiaomi hired robotics lead Kong Tao and his team from ByteDance. ByteDance absorbed many people from Xiaomi’s robotics team into its Seed team. ByteDance also hired Guo Daya from DeepSeek. DeepSeek recruited about 70 people from ByteDance’s Seed team over the past year. Headhunters say a multimodal algorithm lead search takes more than twice as long as a regular technical role.

Against that market, Xiaomi giving its highest job level to its large-model lead sends a clear message: deliver in large models and you can reach the top of the ladder. That message travels further than a recruitment ad.

6. The level is recognition. The product will decide whether it holds.

When Lei Jun recruited Luo Fuli, reports said he offered a package in the tens of millions of yuan. Outsiders asked whether a large-model lead was worth it. MiMo’s iteration speed and metric gains suggest Xiaomi’s internal answer is yes.

But with Level 22 maxed out, the real test starts now.

Lu Weibing said on the earnings call that “AI is the biggest incremental opportunity in the smartphone industry,” while the whole industry faces a transition. Xiaomi’s AI spending is not defensive. It is offensive. Phones, cars, IoT, and robotics will all depend on MiMo for their AI experience. Model iteration speed, cost control, and on-device adaptation all have to hold up.

Longer term, Xiaomi views AI as a frontier of “physical AI.” Lu Weibing said Xiaomi believes physical AI has broader prospects. That means MiMo’s capability boundary will eventually need to extend from language models to embodied intelligence, from on-screen interaction to perception and action in the real world. Luo Fuli’s joining statement, “intelligence will eventually move from language to the physical world,” points in that direction.

The organization has given its recognition. The product will decide whether it holds. Whether MiMo becomes a long-term, stable AI foundation for Xiaomi’s Human-Car-Home strategy, and whether Luo Fuli can lead a team with an average age of 25 to keep shipping models that matter, are not answered by the Level 22 notice. They are answered by the next model release and the AI experience in the next generation of Xiaomi devices.