Luo Fuli on Agents, MiMo-V2 and AI Research Teams

Zhang Xiaojun interviews Xiaomi AI researcher Luo Fuli about her spring 2026 observations on agents, post-training, model development and organizations. Product assessments and forecasts are her views at the time.

Luo recalls using OpenClaw and a strong model for everyday work. Through repeated interaction, she says, the system can organize her ideas into workflows and reusable skills.

She says she asks an agent about management and screening and gradually develops a workflow reflecting her preferences. She describes the experience as a digital counterpart for a class of tasks.

She distinguishes the general usefulness of coding from reliability in non-code tasks: a model and framework strong at programming will not automatically achieve the same completion quality everywhere.

She describes using a strong model to modify an agent framework and then trying other models. In her view, both the underlying model and framework design affect the resulting experience.

Once basic tasks could be completed, she began focusing on cost and speed. She says extensive experimentation around the Lunar New Year changed her view of products and research.

She discusses many people modifying an open agent framework and argues that model capability and a collaborative environment can bring together contributions from different backgrounds.

Luo argues that specialized training can improve agent stability on particular tasks, while a leading system must also be reliable across a broader range of situations.

Drawing on her own experience and that of people around her, she discusses productivity change, developer communities and organizational opportunities, rather than claiming a statistical conclusion across industries.

She discusses teaching agents organizational conventions through examples and interaction, then turning the experience into skill files. Whether a system follows those conventions correctly still needs testing.

On privacy and personal data, she is beginning to consider hybrid local and cloud architectures and sees open ecosystems as a way for more participants to explore them. She stresses that the ideas are preliminary.

Facing rapid change, she says precise predictions a year ahead are difficult. She is excited by recent progress but acknowledges that tools and practices can change over short periods.

On agent iteration, she distinguishes systems helping improve tools from the stronger claim of autonomous continual model improvement, and discusses the roles of people, tools and feedback.

She says multiple agents may improve speed or cost, but she had not seen evidence that they necessarily raise the ceiling of final task capability. This is her observation at the time.

She emphasizes that research requires many functions and team collaboration rather than a single hero, and that tools can help turn experience into shareable processes.

The discussion turns to MiMo-V2. Luo reviews the team’s model goals, product positioning and competitive pressure; performance comparisons are the team’s and guest’s assessments at the time.

She explains that a model intended for reasoning, coding and agents must support multiple uses; quality cannot be assessed from a single benchmark or demonstration alone.

She describes tradeoffs between hybrid attention, long context and inference efficiency. Specific ratios and benefits belong to the model discussed and should not be generalized to all models.

The team also considers multi-token prediction and the balance between computation and memory access to improve practical inference efficiency. She stresses evaluating architecture together with deployment conditions.

Luo says that when post-training cycles grow longer, early assumptions about use cases and context length may become obsolete, so goals and experiments need continual adjustment.

She discusses model pricing, cost and practical user experience and says the team adjusted its pricing approach. Product details and prices may change over time.

She revisits early architecture choices, emphasizing long-context modeling and inference efficiency as constraints while other capabilities require subsequent training and systems work.

On compute allocation, she offers a proposed balance among pretraining, post-training and research experiments. She presents it as a planning view, not a formula for every lab.

She argues that a model-size threshold alone does not determine competitiveness; data, training methods, inference cost, team execution and product experience matter too.

On organization, she emphasizes allowing participants to identify problems, propose experiments and collaborate directly, with managers building culture by solving problems alongside the team.

She discusses reducing rigid hierarchy and barriers between groups so information and compute needs can move faster; such an approach still requires clear responsibility.

Returning to architecture, she discusses parameter growth, the number of attention layers and long-context cost. A larger model does not necessarily improve efficiency on every dimension.

She describes exploring alternative architectures and data reconstruction, with tradeoffs among quality, additional compute and research time.

On multimodality, she distinguishes adding perceptual inputs from improving intelligence and speculates that generation might play a different role. She explicitly calls this a conjecture.

She compares the environments of AI systems and biological evolution, noting access to human knowledge and external resources while acknowledging that their future paths are not directly comparable.

She predicts continued rapid changes in agent frameworks, models and inference demand. These were spring 2026 judgments; the actual pace and outcomes remain uncertain.

On management, Luo says organizational design is complex and requires experimentation, and that changing technical capabilities can alter how a team divides work.

She and the host discuss spending more time on meaningful projects if repetitive work is automated. How benefits and effects are distributed remains to be seen.

She expresses a wish to support basic research, arguing that work without an immediate product application can still deserve time and resources.

They review different approaches to open model ecosystems: some teams emphasize research capability while others focus on broader adoption. She sees value in both objectives.

She reflects on agent evaluations and datasets that may be too specific to a framework, allowing strong test performance without representing real-world task performance.

On competition, she argues that model capability, agent frameworks, resources and ecosystems must improve together rather than optimizing a single score.

She discusses chip supply, compute resources and architecture compatibility, saying teams need to consider feasibility across hardware conditions in advance.

The conversation returns to environment and experience. She says frequent interaction with models, maintaining test cases and probing capability boundaries can help teams adapt to agent development.

On hiring, she values curiosity, fundamentals and diversity and wants an environment where people exchange ideas freely. Her views about experience levels come from her team’s practice.

She argues that agent reinforcement-learning infrastructure must tolerate uncertainty, including interrupted runs, differences between training and inference environments, and scheduling CPUs, GPUs and storage.

Asked for a long-term forecast, she says she prefers to focus on the research immediately ahead. The work and sleep habits she describes are personal, not representative of her team or a recommendation for others.