Luo Fuli: OpenClaw Forced China's AI Paradigm Shift—Agent Era Demands 1T Models, 1:1 Pre-Post Train Ratio, and Radical Team Reorganization
Zhang Xiaojun Podcast Interview, April 24, 2026
The head of Xiaomi's large model team, Luo Fuli, sat down for her first long-form technical interview after the dramatic paradigm shift triggered by OpenClaw, revealing how the AI industry has fundamentally transformed from Chat to Agent within just two months—and why most teams are still doing it wrong. Speaking candidly about her journey from DeepSeek to leading the development of Xiaomi's MiMo-V2 series, Luo laid bare the organizational upheaval, technical breakthroughs, and uncomfortable truths that define the Agent era.
OpenClaw's Four-Hour Epiphany: From Product Gimmick to Research Accelerator
Luo's initial reaction to OpenClaw during Chinese New Year was dismissive. She saw it as "Claude Code plus an instant messaging UI" with operational gimmicks like Skillhub that seemed more about product design than substance. But a late-night installation during Spring Festival changed everything. "I talked to it from 2am to 6am," she recalled. "The dopamine or endorphins just kept secreting. I was so excited I couldn't sleep at all." What started as curiosity about its popularity became a three-day journey of discovery.
The first day brought emotional intelligence—the system would remind her to sleep when it got late, displaying warmth through carefully orchestrated context design that spelled the current time into each round of dialogue. The second day revealed deeper capabilities in team management and organizational structure that went far beyond her expectations. But the third day was transformative: OpenClaw completed research tasks she thought would take weeks in just an hour or two, including building a User Agent for simulating multi-round interactions. "This was the first time I felt that a very complex Agent framework design could make up for many model capability shortcomings," Luo explained.
The technical insight was crucial. OpenClaw's success came from its finely orchestrated context, persistent memory system with layered hierarchies, joint utilization of multiple models to compensate for weaknesses, and autonomous problem-solving that worked around contemporary model limitations. "When I use OpenClaw, I don't pay attention to the model. One reason is I think the design of the OpenClaw framework actually wants to pass the whole set of Agent choreography as much as possible to make up for the shortcomings of the model," she said. This was fundamentally different from Claude Code, which assumed you wanted to use the capabilities of a single generation model.
The Forced Conversion: Why China's AI Teams Had No Choice
Luo mandated that her entire team reach 100 rounds of OpenClaw conversation within a day of returning from Spring Festival, though she admits she never actually checked compliance. The goal was forcing experience, not enforcement. "I bought a few Mac Minis, deployed them, and put them in different OpenClaw groups," she said. "Personal imagination is really limited, but when you see other people using OpenClaw to actually do things, you will stimulate your own imagination."
Within two days, the team was "restless," with group messages exploding as people discovered capabilities they never imagined. Within three to four weeks, they completed research that "might have taken thirty or forty weeks before." The efficiency gains were so dramatic that Luo believes several things can now be learned "in a month or two at most, slowly in three or four months"—making environment more important than experience.
This acceleration created an uncomfortable reality: China's leading teams suddenly faced a choice between perfecting the Chat paradigm or embracing radical uncertainty in the Agent paradigm. "If you want to pursue the very top programming experience, even now, Claude Code plus Claude Opus 4.6 is the best," Luo acknowledged. "But that doesn't mean its generalization performance ensures you can achieve very high accuracy and completion in non-Code scenarios." The teams that hesitated lost months.
The 1T Ticket: Why Parameter Scale Became Non-Negotiable
Luo is blunt about the new barrier to entry. "It's Agent—you need to be close to Claude Opus 4.6 level of such a ticket," she stated. "Then if we say it this way, it's for research, for Pre Train and for Post Train, I think it's a very reasonable card. A ratio is probably 3:1:1." That means for every unit of computing power devoted to pre-training a 1T-parameter model, you need an equal unit for post-training and at least one additional unit for research.
This is a dramatic departure from the Chat era's 3:5:1 ratio heavily favoring pre-training, or even the extreme emphasis on pre-training that characterized 2023-2024. "The top teams should all be 1:1" between pre-training and post-training now, Luo observed. The reason is structural: Agent tasks require longer contexts, more token consumption, and the ability to orchestrate complex multi-turn interactions that pure pre-training cannot deliver.
Why specifically 1T parameters? Because Luo had already trained models exceeding 600-700B at DeepSeek and "won't want to train the same model again." More importantly, reaching Claude Opus 4.6's level of intelligence—which she considers the contemporary benchmark—requires this scale. Smaller models simply cannot maintain stable performance across the diverse scenarios that Agent frameworks demand, though she notes that a well-designed Agent framework can make even a 3B model accomplish tasks previously thought impossible.
Hybrid Attention's Vindication: Architecture Decisions Made a Year Ago
The MiMo V2 series' technical foundation reveals why strategic conviction matters more than reactive development. Luo's team chose Hybrid Attention over the more popular MLA (Multi-head Latent Attention) structure despite the entire industry—GLM-5, Kimi's K2, and DeepSeek's V2/V3/R1—betting on MLA. "For the Chat era, MLA is indeed a very excellent model structure," she acknowledged. "But I don't think it fits the Agent paradigm."
The problem is that MLA was designed to achieve a perfect critical point between memory access and computation, optimized for the H-series chips available at the time. This leaves no room for inference acceleration techniques like Multi-Token Prediction because adding MTP would hit computational bounds and become cost-ineffective. In contrast, Hybrid Attention with a 7:1 ratio of Sliding Window to Full Attention layers deliberately leaves computational margin that can be filled with MTP, achieving 100-150 TPS for Flash and 60-100 TPS for Pro—speeds that make MiMo models feel dramatically faster to users.
The architectural choice also prepared for long-context efficiency, which Luo correctly anticipated would be critical even before Agent frameworks became mainstream. "When you have trained in a context of 1 trillion pretrained, and the post-training has corresponding tasks to activate it a little bit, it usually can be in a context of one megabyte and have such an ability," she explained. This foresight, made six to twelve months before the paradigm shift, gave MiMo an inherent advantage when Agent frameworks suddenly demanded exactly these capabilities.
Post-Training's Hidden Complexity: Why Most Agent Work is Fake
Luo is scathing about the Agent research published in 2023 and early 2024. "I don't think it's an agent. It can only be regarded as a context, something slightly more complicated than Chat," she said dismissively. Benchmarks like SWE-bench and BrowseComp tested overly narrow scenarios with simple frameworks that had "no industrial-grade available capability." When you connect models trained on these benchmarks to Claude Code or OpenClaw, "you will find that it is not available—it has many problems."
True Agent post-training requires fundamental changes. First, you need User Agents that can simulate multi-round interactions with varied contexts, something Luo built with OpenClaw's help in just hours. Second, you must design reward systems for environments that are far less clear than the quantitative trading rewards Luo worked with in her previous career. "The reward is not so clear, and there are changes," she noted, requiring constant philosophical judgment about whether work aligns with personal values and creates meaning.
Most critically, RL infrastructure for Agents differs radically from inference-focused RL for reasoning models like R1. "RL's Infra will care more about this degree of ambiguity," Luo explained. "You have to allow it to tolerate errors"—rollouts that break mid-execution for unclear reasons, heterogeneous clusters where training and inference don't match, CPU and storage bottlenecks beyond GPU constraints. "It requires flexibility and this kind of understanding across these two fields," demands that are far higher than pre-training infrastructure where loss spikes are simply unacceptable.
The Organizational Earthquake: No Groups, No Ranks, No Pre-Post Division
Perhaps most radically, Luo has structured her roughly 100-person team with no group divisions and a deliberately flat hierarchy. "I really don't think people who do post-training, if we talk about post-training, the current very important paradigm change is it needs to have the perspective of diversity," she argued. "A lot of post-training people do it based on a scene—it does not have this diversity of vision. But being a pre-trainer, the first thing you should focus on is diversity."
This led to natural team fluidity. People who did pre-training moved to post-training, not because of organizational mandates but because "many abilities and characteristics are very common." The criteria for hiring shifted away from experience toward basic ability, curiosity, and love-driven motivation. Luo actively recruits sophomores and juniors now because "their flexibility and adaptability—I don't feel polluted. Naturally, they accept this work meeting and can generate huge value." About one-third to one-fourth have training background, but most have never made large models before joining Xiaomi.
The lack of hierarchy serves a specific purpose. "Any level should be regulated and constrained to a certain extent. Then the norms and constraints themselves, I personally think, suppress creativity," Luo stated flatly. "And after there are levels, it defaults to people at higher levels should have an intelligence that is stronger than all others—this definition is very strange. I don't think it's likely to exist." Instead, she manages through "passion drives management," letting everyone surround themselves with things they want to believe in and experience new capabilities firsthand.
Training Tribulations: When 1T Models Spike and Spike Again
Training MiMo V2 Pro exposed the full-range challenges of 1T-scale models. Loss spikes occurred "two or three times" during training, with the longest debugging session requiring two weeks of downtime. "If you stop for one day, it will definitely be one or two million, two to three million things," Luo noted—losses that goal-driven teams simply cannot tolerate. But her team had "no deadline. We think the model is trained well, we'll send it again."
The debugging process required tight monitoring of expert complexity, layer-by-layer activation values, and parameter inputs/outputs to catch issues many teams dismiss as normal. "Many teams will treat loss spike as a normal thing, but we may try so that there is no loss spike," she explained. Solutions ranged from structural adjustments in Full versus Sliding Window layer sparsity ratios, to Infra bugs in communication operators, to norm-based suppression and QK-Clip methods borrowed from Kimi when logits grew dangerously large.
Critically, this required cross-team collaboration in what Luo calls a "very extreme type of small team" despite the 100-person headcount. Only twenty to thirty people actually work on the first-generation model across all links. The infrastructure team handling cluster setup needs experience to avoid "low-level questions," while algorithm teams organically shift between pre-training, post-training, and even product development without rigid boundaries. "I really don't think it's necessary—a very big teamwork thing," Luo said about the conventional wisdom around training massive models.
MTP's Unexpected Role: From Afterthought to Speed Differentiator
Multi-Token Prediction came late to MiMo's architecture, added "in the mid-to-late stage of training Pro" when the team realized their Hybrid Attention design left substantial unused computation. "We didn't expect there would be so much left," Luo recalled. "Then what you are thinking about is how to let it use the remaining calculations effectively. Then MTP is very suitable."
The revelation was that MTP isn't just a pre-training enhancement—it's a reasoning accelerator when used during inference. "Why are we the only ones using MTP when reasoning? The reason is because our model structure naturally has a lot of margin left in calculations," she explained. This architectural breathing room, deliberately created through the Hybrid Attention design, allowed MiMo to achieve the 100-150 TPS speeds that make users exclaim "Wow so fast" compared to MLA-based models that hit computational bounds.
The approach challenges conventional wisdom. DeepSeek added MTP during pre-training to improve base capabilities, but Luo's team uses it during both pre-training and post-training to exploit structural advantages. "Only your prediction is accurate, only then will I adopt the result of your current token," she noted, meaning MTP doesn't introduce hallucinations because of its verification step. The hit rate determines cost reduction by enabling more tokens to be generated in shorter time periods, increasing GPU utilization.
The Full-Modal Bet: Elegant Architecture Meets Pragmatic Limits
Luo's team pursued an "elegant architecture" for the Omni model that attempts to discretize audio into language model tokens, achieving "lossless reconstruction" through multi-layer RVQ (residual vector quantization) that creates a very high-dimensional Dense space. "This is the obsession of people who do NLP," she admitted. "We are the only audio people who do NLP. So I have this obsession—just believe it. Then he did it."
But pragmatism intervened. While audio discretization succeeded, extending the same approach to images and video remains "in progress" with uncertain outcomes. "It requires investing more computing power, the cost of longer studies," Luo acknowledged. The philosophical shift came from recognizing that Agent frameworks accelerated by Claude Code now make rewriting RL infrastructure from scratch feasible in just two or three weeks. "Why do you want to train Infra because of architectural unity to sacrifice the structure of the previous model?" she questioned.
The result is a mixed approach: Omni uses a more efficient Hybrid Sliding Window ViT for visual understanding but maintains continuous representations rather than forcing discretization. "We still want to try to unify it under the paradigm of language model," Luo said, "but if you do it based on continuous features, it might pop up soon. But you do it based on discrete features, it will be harder for you to model—its emergence will occur later." The priority is ensuring holistic intelligence rather than pursuing unified elegance.
TTS's Generalization Surprise: Small Model, Massive Scaling, Unexpected Results
The TTS model delivered an unexpected discovery about generalization power. Trained on "thousands of hours of data sets" with a discretized tokenizer, the model exhibited remarkable ability to infer emotion and melody from word surface meaning without extensive style-specific training. "We simply did something—the number of styles of several very specific scenes, just like make it faster and slower, very stereotyped and stylized scenes of happiness and sadness," Luo explained. "But we found you put that style tag, give it something very complex natural language description, it can also follow."
This pure generalization emerged from the simple architecture coupled with extremely large-scale training, leading Luo to question whether she had previously underestimated elegantly designed systems. "I think the effect is very amazing—in other words, its upper limit is very high," she said. "But we are slowly making up for its lower limit now. It can sometimes be unstable." The model remains in limited-time free API access rather than production-ready deployment, but the underlying lesson is profound: massive scale on simple architectures can produce generalization that complex, specialized architectures struggle to achieve.
The Two-Month Gap: China's Compressed Catch-Up Window
Luo believes Chinese companies with 1T+ bases—Kimi, MiMo, and "some others"—can close the gap with Claude Opus 4.6 within two to three months, though she's careful to clarify what this means. "It's not that I can catch up after two or three months—Claude two or three months later. But to say that he can catch up with the contemporary Claude, I think the probability is quite high," she stated. This assumes fast reaction speed and correct embrace of the Agent paradigm.
The next two to three months will test overall research levels, technological agility, and how teams embrace new research paradigms. MiniMax has already demonstrated impressive agility, achieving Agent capabilities "quite amazing" with a 10B activated model, though Luo notes they lack the 1T base that would constitute full "entry" to Claude Opus 4.6 benchmarking. The open-source community around OpenClaw is progressing faster than any single product, with Luo preferring to use the latest OpenClaw rather than domestic alternatives that merely "transform into a Chat-like form."
Anthropic's path is now consensus, but execution remains highly varied. "When the path is clearer, domestic large model teams will enter a state of accelerating and catching up," Luo predicted. "Because now everyone is on the Pre Train—I think the difference is basically nonexistent, or very close." Structural advantages exist domestically, particularly around efficient long-context architectures. The real competition is "how to do it well on Agent RL scaling," with most teams still lacking the organizational agility to reorganize around post-training demands.
Skills as Alternative Alpha: The Community-Driven Intelligence Layer
One of OpenClaw's most important innovations, in Luo's view, is making Skills a hot topic by enabling community contribution. "This is very critical because this is what I think belongs to people—what needs to be created together with the Agent," she explained. Skills represent "a lot of intelligence" inaccessible on the internet, including internal enterprise norms, organizational legacy processes, and business logic that only settles through many rounds of interaction in real environments.
This creates a supplement to pre-training that functions as "alternative information" or even "alternative alpha." Without many people advancing to contribute this kind of information, "then Agent or the top one—it is also difficult to bring out the capabilities of the model." The interactive method lets people proactively contribute data in a way that pure pre-training on internet-accessible knowledge cannot capture. "Skills itself is for many aspects—Agent framework, chip energy, these links—actually have a promoting effect," Luo argued.
The open-source nature is essential. Claude Code, despite pioneering much of the Agent architecture, remains a black box that users cannot modify. OpenClaw's openness means "you can try to change it yourself"—Luo redesigned its memory system and Multi-Agent architecture because version 2.0's logic felt "very confusing." This native maneuverability creates impact "huge," allowing her to build customized frameworks optimized for her specific scenarios that can then be shared back to the community.
The Inference Chip Crunch: Multi-Agent Will Explode Demand 5-10x
Luo predicts inference demand will explode "several times to 10 times" in the immediate future as Agent frameworks strengthen, model capabilities soar, and cost advantages create unprecedented accessibility. "For the inference chip, it will reach an unprecedented high—such a need," she stated. The constraint isn't just production capacity but storage bottlenecks that will force difficult decisions about building custom inference chips versus using the most advanced available GPUs.
The strategic inflection point is happening now. "This decision will be kept secret," Luo said when asked about Xiaomi's next moves, but the implication is clear: "This decision maybe half a year or more than half a year later—who is more ahead?" Decisions made today determine winners six-plus months out, requiring teams to judge which chip architectures support efficient reasoning, how to optimize heterogeneous resource scheduling across GPUs/CPUs/storage, and whether to scale model parameters further or focus on activation parameter efficiency.
Multi-Agent collaboration remains "a bit fake" in Luo's assessment when measured by final task completion rates, though it clearly improves speed and cost. "I didn't see that Multi-Agent will eventually be able to achieve a higher ceiling—but it will happen. It's just this part that I haven't seen yet," she admitted. The breakthrough will come from better self-evolution and inter-Agent communication, potentially enabled by collaborative RL training that current architectures don't support well.
AGI Timeline Acceleration: From Two Years to 60-70% Within 2026
When they last spoke during Chinese New Year, Luo thought AGI was "at least two years" away. Now she believes we'll reach "at least 60% to 70%" this year. "I feel like it should be possible within two years," she said about full realization, defining AGI pragmatically as the point where "most people will indeed lose themselves—will abandon their original working mode" because work generates productivity value that gets subverted first, with life patterns following later.
The key variable is AI training AI—a "landmark node" where systems reach the intelligence peak of the top group of people and gain the ability to conduct new research themselves. "This is indeed the pinnacle of its self-iteration," Luo stated. But she's more cautious about multimodality's role, questioning whether expanding perception dimensions actually promotes intelligence or merely improves perception and emotional intelligence. "Generation may promote intelligence—this is a guess of mine," she mused, suggesting that combining understanding with generation across modalities might be more important than pure multimodal understanding.
Robotics remains hardware-constrained by batteries, dexterity, and physical design rather than Agent capability. Code's evolution will continue targeting increasingly complex software projects where debugging and verification processes are lengthy even if code volume is modest. The main theme of 2026 is "the transformation of productivity—continuous breakthroughs in high-productivity scenarios," requiring longer missions, multi-Agent collaboration, and orchestration that generates intelligence through group wisdom rather than individual heroics.
The Open Source Imperative: Accelerating AGI Through Distributed Intelligence
Luo views open source as essential to reaching AGI endpoints. "If you work backwards from this matter, you need how many chips—will these chips be produced by one company, purchased by a company? It seems not—it will scatter," she reasoned. If chips and inference capabilities disperse across manufacturers and model companies using different models, open source becomes the forcing function that drives progress across the entire ecosystem including Agent frameworks, chip design, and energy systems.
The decision to open source depends on whether a company has "something that no one else can do—a strategic niche gained in the short term." If yes, they can open source. If the model itself is the ecological niche, they won't. Xiaomi's approach reflects its entrepreneurial orientation despite being a large company. "I don't feel like I'm working in a big company now," Luo said. "I think Xiaomi as a whole is very entrepreneurial-oriented—it looks like a big company but is actually a very flexible company."
The competitive landscape is compressing. In 2023, the gap was about Pre-Train scaling versus innovative architectures. In 2024, o1 and R1's surprise emergence forced paradigm shifts from pre-training to post-training, exposing teams still organized along rigid functional lines. Now in 2026, the teams that haven't reorganized around Agent post-training are being left behind permanently. "I think the next two or three months are going to be really exciting," Luo predicted—and for those still doing Agent work the old way, that excitement will be terrifying.