Luo Fuli on Agent Frameworks, Model Training, and Teamwork

Zhang Xiaojun introduces Luo Fuli, head of Xiaomi’s large-model team, for a long technical interview recorded after the releases of OpenClaw and Xiaomi’s MiMo V2 models. They discuss agent frameworks, model training, organizational change, and possible future directions. The video identifies its English translation as AI-generated and intended for reference.

Luo describes OpenClaw as a turning point in how she thinks about agents. At first she saw it mainly as a programming agent connected to messaging, with a new interface and an always-on experience. Trying it herself changed her assessment.

During the Spring Festival she installed the tool late at night and stayed up talking to it. Its attention to context, including awareness of the current time, gave her an impression of warmth and autonomy. She attributes that impression in part to the way the framework organizes context.

She next tried tasks she had doubted the framework could handle, including discussions about team building, curiosity, hiring, and organizational structure. She says the system helped organize those ideas into reusable skills; this is her account of her own use, not a controlled evaluation.

On a third day she explored research tasks, including building a user agent to simulate multi-turn interactions. She says the result could help create richer agent scenarios for supervised fine-tuning and reinforcement learning, and describes the experience as a shift from product novelty toward research assistance.

Luo points to persistent memory, model coordination, and autonomy as parts of the framework that work together. In her experience, an agent can sometimes compensate for a model’s limitations by finding a suitable tool or another model, though the outcome depends on both the framework and the underlying capabilities.

She says her team connected MiMo V2 Flash and a smaller model to the framework and observed useful behavior without extensive task-specific training. She presents this as a reason to study how orchestration can improve the performance of less capable models, not as a universal benchmark result.

The resulting research question is how to make models work consistently across different agent frameworks. Luo argues that post-training must adapt to agent tasks rather than only chat, including different tools, contexts, interaction patterns, and environments.

She distinguishes serious coding from broader everyday tasks. A coding-focused agent can be strong at software engineering while a general-purpose framework adds features such as messaging, scheduled work, and persistent memory that support different forms of interaction.

In her account, framework design is a middle layer between the user interface and the model. It governs context, memory, tool choice, and orchestration; the visible interface is only one part. She expects coding and general agent products to borrow useful design ideas from each other.

Luo says a good framework can help a mid-level model handle many tasks, while difficult long-horizon or demanding programming work may still require a stronger model. Her illustrative percentages are personal estimates, not independently measured performance figures.

She credits Claude Opus 4.6 with the strongest experiences in her initial tests and says she used it to modify an open-source framework. Once she had improved the setup, she could try other models as well. Her larger point is that model and framework development can reinforce one another.

For Luo, agent self-improvement includes adapting both the model and its surrounding system: stored memories and skills, as well as the dynamic workflow that decides what information and tools to use. Different applications may need different architectures and tradeoffs.

She values open-source frameworks because users can inspect and modify them for their own circumstances. She contrasts that flexibility with a closed system where internal memory and workflow choices cannot be changed directly.

Her early experiments were costly: she recalls spending roughly $1,000 on the first day and testing lower-cost alternatives. As her setup improved, she began focusing on what tasks remained possible and how to reduce cost and latency. The expenditure is her recollection of an individual experiment.

After the holiday she encouraged colleagues to try the system, setting up machines and group conversations so people could share discoveries. She says her forceful message about using it was meant to convey urgency; she did not actually check or dismiss people based on a conversation count.

The shared experiments helped colleagues see one another’s uses and improve the framework together. Luo says the group’s iteration was faster than working alone and helped the team explore how its own model could operate in an agent environment.

Luo notes that broader testing still revealed reliability gaps compared with Claude. Nonetheless, she believes a community of users can discover problems and improve an agent framework much faster through shared experiments.

She says an agent can accelerate parts of research that once took days or weeks, allowing multiple ideas to be tested in parallel. In her view, choosing worthwhile questions and designing evaluations become more important when implementation gets faster; her speed comparisons are personal estimates.

Luo connects coding with long-horizon tasks. She argues that code repositories contain structured dependencies across files, which may provide useful training material for handling long contexts. She distinguishes that hypothesis from a guarantee that coding ability transfers reliably to every domain.

Software development also requires planning, revisiting earlier changes, and managing context, so agent frameworks developed for coding can offer ideas for other long tasks. Her team wants to train on varied tasks using both supervised fine-tuning and reinforcement learning.

To broaden coverage, she describes having many people use the model and then identifying additional scenarios. A difficult part is recreating an environment that supports multi-step interaction and defining useful reward signals. She does not disclose the team’s internal implementation.

She notes the practical expense of training on very long trajectories: generating and evaluating them takes time and compute. She describes a combination of long-context pretraining and shorter post-training tasks as an approach under consideration, while cautioning that capabilities and costs change quickly.

Asked why OpenClaw appeared especially popular in China, Luo offers a personal view: demand for productivity tools and access to relatively inexpensive models may encourage experimentation. She says she does not have a complete view of adoption in Silicon Valley.

She distinguishes earlier narrow agent evaluations from the demands of a practical agent. A high score on a bug-fixing benchmark may not capture whether a model can understand a framework, manage context, and complete a broader software or business workflow.

Skills can represent procedures and organizational knowledge absent from public pretraining data, she says. People may contribute that context through repeated interaction, and agents can help organize it into reusable instructions. She regards this as a collaborative process rather than a replacement for all model knowledge.

Luo expects the impact of agents to spread over time across many less clearly defined tasks, unlike a benchmark with a single ground-truth answer. She stresses that evaluation remains necessary as systems improve and differences become harder to judge by experience alone.

On privacy, Luo imagines a hybrid system in which smaller local models handle some sensitive tasks on a device while more demanding work goes to the cloud. She says both models and frameworks need further development; local processing alone does not resolve every privacy or security issue.

She sees two open questions beyond training models for agents: whether frameworks can improve themselves and how people and agents can learn together. At the time of the interview, she had not seen those ideas realized broadly.

Luo describes two goals that can require different tradeoffs: pushing performance on high-value, long-horizon tasks and making useful assistance accessible across everyday situations. In the latter case she emphasizes multimodal input, cost, and speed.

She believes broader participation is changing AI work: developers and users can contribute skills, framework changes, and research ideas rather than leaving every improvement to model researchers. That is her explanation for the excitement around agent tools.

Even as models improve, she expects orchestration to matter for cost and speed. Smaller models may eventually deliver more capability at lower cost, though the exact timing and performance are predictions rather than established results.

On multi-agent systems, she says parallel work can improve speed or cost but cautions that she has not yet seen clear evidence that adding agents always raises the final quality ceiling. She expects future research to test whether collaboration can do more.

Her longer-term goal is for agents to take on increasingly complex research and management work, including data access, computation, and evaluation. She contrasts this research objective with optimizing for daily active users; these are her priorities, not a claim that other goals are invalid.

She discusses the possibility that agents could help researchers develop stronger models, while acknowledging that this is a future-oriented hypothesis. Her analogy to team work emphasizes specialized agents with different contexts coordinating on a larger task.

A capable framework, in her view, supplies information a model would otherwise miss through memory, messaging, proactive tasks, and tool coordination. It also needs evaluation: currently people often provide feedback and correction, and better automatic assessment remains a challenge.

Turning to MiMo V2 Pro, Omni, and TTS, Luo says the team wanted language reasoning, multimodal perception, and speech output to work together within agent workflows. She frames their performance as a team goal and personal assessment, not proof that the combination matches human intelligence.

Keeping specialized models separate can reflect cost, speed, and task requirements: a speech generator need not always use the largest language model. Orchestration should choose capabilities according to the task, while coordination and multimodal understanding still need improvement.

She says MiMo V2’s design emphasized long-context efficiency, including speed and cost. She discusses hybrid attention and contrasts it with other architectures, explaining that hardware bottlenecks and inference design affect tradeoffs. The technical and performance claims are her account at the time of recording.

Luo describes combining sliding-window and full attention with multi-token prediction to balance long-context efficiency and inference speed. She says the team found useful compute headroom during inference design; the best combination depends on architecture, hardware, and workload.

She argues that model design needs to account for deployment hardware and post-training, whose duration and requirements may change. An architecture optimized for one known setting may be efficient there, while a simpler design may leave more room to adapt to future tasks.

Her team experimented with the balance between full and sliding-window attention at different model sizes. She presents its sparsity observations as experimental results rather than a universal rule, and notes that other teams may obtain different results.

She discusses pricing as another changing choice: architecture and inference cost matter, but post-training and the value a model creates in an agent workflow may also affect pricing. Her comments describe the team’s reasoning rather than a fixed rule for the market.

Luo says MiMo V2 Flash and Pro were trained in parallel. The larger model required more work on training stability, including investigating numerical spikes and uneven expert activation. Debugging involved infrastructure and research teams working together.

She discusses compute allocation across experiments, pretraining, post-training, and inference. Her suggested research-to-pretraining-to-post-training ratio is an illustrative planning view from the interview, not a general formula; actual inference needs depend on users and tasks.

For large-model training, Luo says the first step is recognizing an instability as a problem. Her team monitors activation values, parameters, and expert behavior, then tests whether architecture, numerical behavior, or infrastructure bugs could be responsible.

Possible mitigations include clipping extreme values or changing normalization, each with tradeoffs. She describes small teams diagnosing problems across infrastructure and research, sometimes pausing a costly run to investigate instead of treating a deadline as the sole goal.

On organization, Luo emphasizes that different specialties need to work together and that broad experimentation can inform research choices. She returns to her earlier example of encouraging colleagues to try an agent, while saying she did not enforce a fixed conversation quota.

Luo says people who worked on pretraining data also moved into post-training, bringing experience with data diversity. She favors flexibility in assignments, while noting that infrastructure, data, training, and product work still require different skills.

She describes a flat team structure in which people lead particular projects without necessarily having fixed hierarchical control over every participant. In her view, reducing rigid divisions can help people contribute ideas and move between stages of training.

Her management philosophy emphasizes curiosity and shared motivation. She believes hands-on experience with a new tool can stimulate interest, though the team still needs coordination, technical foundations, and a way to solve practical problems.

Training runs can suffer loss spikes, she says. Her team chose to stop and investigate recurring instability, sometimes spending days or longer testing possible causes. Even without a fixed release date, repeated failures and the use of scarce compute created pressure.

She argues that parameter scale and context both affect capability, but also recognizes the tradeoff between activated parameters and inference cost. Her suggested scale for reaching a particular agent capability is a personal assessment at that time, not a proven minimum.

Long-context training requires useful, dense supervision across the context, which can be difficult or expensive to construct. She describes gradual improvement through data and training choices rather than assuming that a large advertised window guarantees effective use of every token.

For multimodal work, the team explored representing audio as discrete tokens and studied the reconstruction and training tradeoffs. Luo says a unified language-model approach is attractive, but emphasizes that unification should not come at the expense of overall performance.

She says coding agents have made it faster to build infrastructure, changing her thinking about whether every modality must use an identical architecture. A team may choose separate implementations when that produces better results instead of pursuing elegance for its own sake.

Discussing MiMo V2 Omni, she describes joint perception across text, images, audio, and video as useful for agent interaction. She reports impressions from use, but says benchmark evidence and the effect of multimodal training on general intelligence remain open questions.

She distinguishes multimodal perception from generation and treats possible benefits of combining them as a hypothesis. For TTS, she describes experiments with discrete audio and style instructions, while acknowledging that the model was not yet consistently production-ready at the time.

Luo compares model development with biological evolution, arguing that their environments and starting conditions differ. She notes that current models rely on human knowledge, compute, and objectives; her claims about future model behavior are speculative.

She expects coding agents to tackle more complex development and verification, not merely produce more lines of code. Beyond screens, she sees robotics as another possible interface, while recognizing that hardware, batteries, and manipulation impose their own constraints.

She gives an aggressive personal estimate for progress toward AGI, while acknowledging that the term lacks a clear definition. Such percentages and timelines are her predictions at the time of the interview, not measured progress or a consensus forecast.

Comparing model teams, she expects rapid changes over the following months and argues that research agility and adapting to agent workflows matter. She distinguishes catching up to a contemporary model from catching up to whatever that model might become later.

Luo expects agent use to increase demand for inference and says chip availability and memory may become constraints. Decisions about model architecture, compute supply, and hardware need to be made ahead of training; her demand projections are uncertain.

She summarizes three qualities she values: long-range architectural and infrastructure planning, agile post-training, and curiosity or commitment. Faster agent-assisted research may shorten planning cycles, but teams still must make bets before outcomes are known.

Agent-focused reinforcement learning infrastructure must coordinate a model with tools and environments, she says, instead of handling only a model’s isolated response. That raises questions of compatibility, fault tolerance, and scheduling across different resources.

She treats group learning as part of the research process: people bring distinct expertise, share results, and refine a framework together. She also discusses wider open-source collaboration as a way for more contributors to inspect and improve systems.

Reflecting on her recent work, Luo says she frequently revises earlier judgments as tools and circumstances change. She wants her work to be useful to others and considers supporting basic research one possible future contribution.

Looking back from ChatGPT through subsequent model releases, she describes how open model teams developed different emphases: broad ecosystem support, model scaling, and architectural efficiency. She sees those approaches as complementary contributions rather than a simple winner and loser.

She recalls the shift toward reasoning and post-training, including work associated with DeepSeek R1. The experience led her to pay more attention to whether a capability developed in code or mathematics might generalize beyond its initial task.

Luo argues that a team needs room to learn as methods change. She describes giving relatively inexperienced colleagues opportunities to build systems and experiment, while avoiding overly detailed instructions that might suppress their own ideas.

For agent training, she stresses more complex, end-to-end tasks and environments instead of optimizing only a narrow benchmark. She says concrete methods for scaling agent reinforcement learning are still under exploration and does not disclose her team’s internal results.

Competition, in her view, involves faster iteration and more interdependent decisions: pretraining, post-training, context length, frameworks, and inference hardware all affect one another. She expects small teams to gain leverage from agents but acknowledges current multi-agent systems still have limitations.

She supports open source as a way for more people to develop and deploy models and frameworks across different hardware. Companies may make different choices depending on their strategy and resources; her claim that open source accelerates progress is an argument, not a guarantee.

Asked what model companies need to compete, Luo points to a strong foundation model, an adaptable agent framework, and the ability to use existing resources effectively. She also asks organizations to reconsider how people and agents can work together as productivity changes.

She describes research as combining original ideas with scalable engineering, citing model architecture and long-context efficiency as examples. Her team values its own experiments and external feedback; she accepts that some views expressed in the interview may later prove wrong.

Luo says she values a learning environment, curiosity, foundational skills, and diversity more than treating prior experience as a fixed measure of potential. She describes recruiting people from different backgrounds and encouraging them to share ideas and learn from one another.

Post-training requires both people who test models closely and infrastructure engineers able to handle ambiguous failures. She distinguishes the precision required in pretraining from the fault tolerance and resource scheduling needed for complex agent rollouts.

She says large-scale agent reinforcement learning and continual interaction remain difficult research problems. At the end of the conversation, she describes her personal work schedule and explicitly says it does not represent the rest of her team.