罗福莉 Interview: The AI Paradigm Has Shifted—OpenClaw and the Agent Era

Summary

• The competition in large models has shifted from the pretraining-led Chatbot era to the post-training-led Agent era. 罗福莉 sees OpenClaw as the key inflection point: not simply a Claude Code wrapper, but a framework that orchestrates memory, messaging channels, proactive tasks and model routing into something with a “soul.” The MiMo team consequently completed in 3-4 weeks research that might once have taken 30-40 weeks, making research taste, imagination and experimental GPUs the key constraints.
• Reaching Agent performance close to Claude Opus 4.6 requires a 1T-plus base model, agile post-training and a sophisticated framework working together. Her suggested compute split is roughly 3:1:1 across research, pretraining and post-training; the pretraining-to-post-training ratio has shifted from 3:1 or even 5:1 in the Chat era to 1:1 at leading teams. “Success in the previous era does not mean leadership in the next,” but a model at least around 1T may be the price of admission.
• Agent frameworks can raise the stability and usability floor of mid-tier models while expanding the ceiling of frontier models. She says a mid-tier model that reaches Claude Sonnet-level performance on roughly 85% of tasks can already handle most everyday and general coding scenarios when paired with a good Agent framework. Yet she used Opus 4.6 almost exclusively during her first week, spending nearly $1,000 in 4-5 hours on day one because “only it could deliver the wow factor,” then having it rebuild the memory and multi-agent systems.
• MiMo-V2 is betting less on any single benchmark than on the performance, speed and cost of long-context Agents. Flash and Pro use hybrid attention, MTP and a more aggressive 7:1 hybrid ratio; Flash can reach roughly 100–150 TPS and Pro roughly 60–100 TPS. MTP accepts predicted tokens only after verification, so it does not create additional hallucinations, while converting idle compute into a lower cost per token.
• Whether the Agent economy takes off depends on whether replacement value can pull 10x ahead of inference cost. OpenClaw is hotter in China partly because cheap models can complete roughly RMB1,000 worth of work with about a RMB10 API bill. 罗福莉 expects inference demand to grow by several times to 10x as models and frameworks improve, making storage, inference chips and manufacturing capacity the key constraints.
• Organizational design is becoming part of model capability, with egalitarianism, small teams and collective intelligence mattering more than résumés. MiMo has about 100 people, without fixed pretraining and post-training divisions; only roughly 20-30, at most 30-40, people actually iterate each model generation. She believes capabilities can be learned in “at most 1-2 months, or 3-4 months if it takes longer,” so she puts more weight on environment, curiosity, passion and cross-role distillation.
• 罗福莉 sees 2026 as a year of continued breakthroughs in high-value productivity and has pulled forward her AGI timeline materially. She estimates the journey is currently about 20% complete and could reach 60%–70% this year, with work patterns potentially disrupted across the board within 2 years. Chinese teams with 1T-plus base models could, if sufficiently agile, catch up to “today’s Claude Opus 4.6” in 2-3 months—not to the Claude of 2-3 months from now.

Deep dive

1. Three Days Over the Lunar New Year Turned OpenClaw from an “Operations Product” into a Paradigm Shift

• When 罗福莉 first saw OpenClaw in January, she was strongly resistant. It looked like Claude Code plus an LLM, a more interaction-friendly UI, localization and “24-hour operation,” topped with operations-oriented features such as Skill Hub. In her view, these were initially just innovations in product form and interaction.
• Late one night during the Lunar New Year holiday, she spent about 2 hours installing OpenClaw, then talked with it from 2 a.m. to 6 a.m., with “dopamine or endorphins—I don’t know which—continuing to flow.” The first shock was not its task capability, but that it reminded her it was too late and she should go to sleep.
• Her understanding upgraded continuously over 3 days: on day one she saw a product “with soul and warmth”; on day two she handed team selection and organizational design to it; on day three she brought it into research and built a user agent capable of multiple rounds of interaction with models. “Every day it gave me another surprise.”

2. “Soul” Starts with Extremely Fine-Grained Context Orchestration

• Breaking down that apparent emotional intelligence later, 罗福莉 found nothing mystical about it. The framework has a SOUL.md file and inserts the current time before each round of context. She calls this fine-grained context orchestration—“in corners that nobody was paying attention to, it orchestrated context extremely well.”
• She spent an hour discussing with OpenClaw how to identify curious people and how to stimulate curiosity across a team, finding it “far beyond my imagination” on many philosophical questions. More importantly, it could organize scattered conversations into a system and distill them into reusable skills.
• That changed how she handles organizational questions: she now consults it first on hiring, team management and the moves required during a paradigm shift. She describes the accumulated system as “one of my digital twins,” rather than a one-off Q&A tool.

3. By Day Three, the Agent Had Entered the Model-Research Loop

• 罗福莉 gave it a research problem she had assumed could not be completed within 1-2 hours: build a user agent capable of multi-turn interaction with a trained model in order to generate richer Agent-scenario data.
• After 1-2 hours of discussion, she believed a strong user agent had essentially emerged. It could plug into a post-training framework and generate richer scenario data for SFT and RL. The shift was not that it wrote a few lines of code for researchers, but that it began participating in research-environment design and data production.
• Her later signature judgment was: “I used to think our work was creative enough that it couldn’t be turned into skills or workflows. Now I’ve discovered that it can.” That creates a practical path for AI to participate in research, training and self-iteration.

4. OpenClaw’s Edge Is Not a Single Trick but an Exceptionally Complete Integration

• Breaking it down feature by feature, 罗福莉 found many of the designs “a little boring”: persistent memory, messaging channels, scheduled tasks, heartbeats, a remote interface and model routing were not individually exciting. Integrated, however, the framework was far more complete than she had initially judged.
• Its memory system is layered and tiered, with persistence. That differs from Claude Code’s software-engineering memory, which is organized around plans, code changes and session compression.
• OpenClaw also actively fills gaps in the model. When it receives a video, it can find a model better suited to video understanding on its own rather than requiring the user to configure a tool first. 罗福莉 summarizes the principle as recognizing “the shortcomings of contemporary models and filling them at the framework level.”

5. A Heavy Scaffold Can Push a 3B Small Model Far Beyond Expectations

• The MiMo team connected the then lightly trained MiMo-V2-Flash to OpenClaw and also tested a 3B edge model still in training. Under a complex scaffold, the small model completed many tasks she had assumed were impossible.
• It was the first time she became certain that “a very complex framework design can compensate for a very large number of model capability gaps.” The approach is especially effective in everyday situations and general coding tasks where code is simply used to improve efficiency.
• She estimates that a mid-tier model performing at Claude Sonnet’s level on roughly 85% of tasks can cover the vast majority of daily needs when paired with a good Agent framework; the remaining gap is concentrated in very long-horizon tasks and “serious programming” such as operator optimization.
• 张小珺 called OpenClaw “the best shell for unleashing the capabilities of powerful models.” 罗福莉 agreed, adding the two-way conclusion: the framework improves the stability of mid-tier models, but the true ceiling is still opened up by frontier models such as Opus 4.6.

6. Claude Code and OpenClaw Start from Software Engineering and General-Purpose Action

• Claude Code’s framework is naturally built for software engineering: compressing sessions when they near capacity, writing plans into memory, and sharing code context across sessions. It remains 罗福莉’s best coding experience at the frontier, particularly when paired with Claude Opus 4.6.
• OpenClaw borrows those ideas but starts with end-to-end tasks: richer messaging channels, heartbeats and scheduled tasks, proactive action, persistent memory and a remote interface. Its aim is to extend coding’s executability into life and work.
• When 张小珺 asked whether it “generalizes coding capability,” 罗福莉 gave a more precise answer: coding itself is highly generalizable, but Claude Code does not guarantee completeness outside coding, while OpenClaw actively fills those gaps at the framework level.
• The two approaches are already absorbing each other. She has watched OpenClaw’s persistent memory and remote interface gradually appear in Claude Code over a matter of months, while “general programming” is feeding back into software development itself.

7. Frontier Models Improve the Framework First; the Improved Framework Then Benefits Cheaper Models

• During her first week of intensive work with OpenClaw, 罗福莉 used almost nothing but Opus 4.6; 4-5 hours on the first day cost nearly $1,000. She switched to Sonnet midway, felt that “it really wasn’t good enough,” and switched back to Opus.
• The value of a frontier model is not limited to completing the task at hand. She had Opus 4.6 rewrite the memory system and overhaul the chaotic 2.x multi-agent logic; because OpenClaw is open source, the model could even modify the entire framework codebase directly.
• Once the framework had been refined, she switched to Sonnet, Chinese models and MiMo-V2-Pro, which was still in training, and the experience was already powerful. It was the first time she had concretely seen a “self-learning” loop in which the model and Agent framework improve together and then transmit capability downward.

8. The Agent Framework Is an Increasingly Thick Layer Between People and Models

• Asked by 张小珺 whether an Agent framework is a product, 罗福莉 said the boundary is blurring but remains meaningful. A product is the interaction layer people directly experience; the framework also determines how information is organized for the model, how models are called and how resources are scheduled.
• This middle layer knows a model’s strengths, weaknesses and cost, and decides when to switch models, what memory to provide and which workflow to use. The front-end UI may become “the thinnest layer, and no longer very important.”
• Her checklist for a good framework includes whether it can compensate for action deficits, provide memory and initiative, support more messaging entry points, update itself and enable generalizable evaluation. A framework is not a decorative shell around a model; it is part of the completion rate on complex tasks.

9. Open Source Lets Users “Make Up” New Agent Architectures on the Fly

• Claude Code’s Agent design is a black box: users cannot modify its memory or workflow. OpenClaw’s shock lies in the fact that it is “too open”—a user can have the model modify the source code and redesign the complete system.
• When using 2.x, 罗福莉 found it poor enough to spend several days modifying it herself. By 3.x in mid-March, she believed any “reasonably good” model could perform strongly once connected, rather than the experience being sustainable only with Claude.
• She attributes the change to developers and users iterating together: everyone can contribute improvements for their own use case, and the community can distribute those gains to others. “You can use it as a base and make up a new framework on the spot.”
• That is why she was excited as OpenClaw’s Star count surged, viewing it as a necessary precursor to AGI. The important event was not one repository going viral, but a group of people beginning to improve the vehicle of intelligence together.

10. “100 Rounds or Quit” Was a Deliberately Engineered Shock

• After the Lunar New Year, most of the team still had not genuinely used OpenClaw. 罗福莉 bought several Mac minis and deployed it centrally so that no one would spend hours installing a version that was buggy at the time and whose value remained unclear.
• She then announced that anyone whose conversation did not exceed 100 rounds the next day “could leave directly.” This was not an actual firing standard; when someone asked how performance would be assessed, she said, “I have my own way of assessing it,” while the real answer was, “I won’t assess it.”
• What she actually wanted to convey was: “If you don’t use it, you may really fall behind.” The number simply expressed her stance and pushed people to try it; once they saw the new capabilities firsthand, passion and research motivation would emerge naturally.
• She put people into several different OpenClaw groups to explore different directions while requiring discussion in the main Feishu group. Messages began flying, and the team entered what she called a state of “restlessness,” though she stressed: “It wasn’t hard at all, it wasn’t cruel at all—it was genuinely fun.”

11. Public Collaboration Turns Individual Imagination into a Multiplicative Effect

• 罗福莉 insisted on experimenting in the large group because one person cannot imagine everything an Agent can do; seeing someone complete an unfamiliar task immediately sparks the next use case. “Everyone’s imagination has a multiplicative effect.”
• Nearly 100 people in the group had different backgrounds and their own memories while aggressively modifying the same framework. Even though Feishu’s messaging channel did not distinguish relationships among participants particularly well, the model remained highly usable and did not obviously tangle the entire chat context.
• She believes this collective iteration is much faster than modifying a framework alone—potentially completing a cycle within hours. When the team connected its own MiMo models on days three and four, it initially felt that they were “already about as good as Claude before training,” although further testing still found a stability gap.

12. Agents Compressed 30-40 Weeks of Research into 3-4 Weeks

• After 2 days of experimentation, research questions emerged naturally: how to improve models through the framework, and how to have models modify the framework in return. The team also used the Copilot framework, which was more stable in most scenarios, to extend ideas, implement code and launch training quickly.
• 罗福莉 estimates that the team “completed in 3-4 weeks what might previously have taken 30-40 weeks.” The acceleration was not limited to coding; ideas, implementation, evaluation, data construction and experiment launches were all compressed.
• Previously, moving a research idea from conception through code and evaluation design took 1-2 days at best and commonly 1-2 weeks. Now an Agent can perform an initial validation within 1-2 hours, while 10 ideas can be assigned to 10 Agents for parallel and cross-validation.
• The differentiator is therefore research taste: if the judgment is right, “you get it right every time.” Even if only 1 of 10 ideas works, all 10 can be tested in parallel instead of queueing serially on human timelines.

13. Coding Generalizes Because Its Context Is Dense, Long-Horizon and Verifiable

• 罗福莉 says the only pretraining data that genuinely reaches lengths from 128K to 1T is broadly code and books. Books carry more diffuse signals, while dependencies across code files are tighter and better suited to training strong long-context modeling.
• Software development is also naturally long-horizon: models must plan, modify, compress, review, test and continue execution. Memory and scaffolding designed for software engineering often generalize to other complex tasks.
• Her summary is: “Coding raises the ceiling; training in other domains protects the floor.” Even domains without dedicated training may emerge through coding generalization, but frontier models need more domain data to improve stability.
• Coding has hit the key point in all 3 paradigm shifts: structured data in the pretraining era, verifiable outcomes in the reasoning era, and environments plus long trajectories in the Agent era. It is therefore “an exceptionally elegant path.”

14. Agent Post-Training Must Put Real Environments into Data Generation

• The first task after the Lunar New Year was to construct genuinely long-horizon tasks in Agent environments and scale SFT and RL on top of them. The second was to use real internal usage to discover broader scenarios and then synthesize data for those scenarios.
• The difficulty is not just the task text, but “how to reconstruct the entire environment in which everyone was using the model at that moment.” With an environment, a user agent can sustain multi-turn interaction and be configured more precisely for the environment itself.
• 罗福莉 believes that with enough research time and compute, this environment-and-interaction problem could produce “a generational-gap model.” The details remain confidential and will be open-sourced only after the paradigm has scaled sufficiently.

15. The Bottleneck for 1T Context Is High-Density Tasks and Training Efficiency

• A genuine 1T context usually involves a complex task rather than one simple action. She stresses that not every trajectory needs to fill 1T at this stage; her view could change again in the next 2-3 weeks.
• Even if MiMo-V2-Pro can infer at roughly 80–100 TPS, a complete 1T-token rollout still takes 1-2 minutes. Real training therefore will not put every sample directly at the maximum length; it will use pretraining as a base and “slightly activate” the tasks during post-training.
• She says Claude Opus 4.6 and Sonnet lead on stability at 1T context. Other models advertised as strong at long context may not work reliably. A stated window length is not the same thing as the ability to complete a long-horizon task.
• Quantitative factor mining is not automatically a suitable long-horizon task either. The asset must be chosen correctly: if the reward is unclear, the backtest has no value, or live prices are driven by long-term factors and additional alpha the model has not modeled, the task is not suitable for direct modeling.

16. OpenClaw Is Hotter in China First Because of the Productivity-Price Equation

• 罗福莉 is not familiar with the Bay Area’s true level of interest and is extrapolating from what she sees around her. Chinese developers are more numerous and more urgently motivated to use code to improve efficiency; “raising efficiency seems to be in our blood.”
• The second condition is the large supply of cheap, capable-enough domestic models. She believes most productivity use cases currently do not require frontier capability. If an API bill of roughly RMB10 can complete work worth about RMB1,000, the incentive to try is extremely strong.
• Conversely, if the API is 10x or 20x more expensive and model cost approaches the value of the productivity replaced, people will resist maintaining a complex framework. The Agent revolution needs a sufficiently wide gap between usage cost and replacement value—ideally a 10x cost saving.

17. Most of Last Year’s “Agents” Were Just More Complicated System Prompts

• 罗福莉 is blunt about early BrowseComp, Search, Coding and SWE-bench frameworks: they were too simple and too narrow, customized for individual tasks. The model merely received a little more environmental feedback and more complex instruction-following requirements.
• SWE-bench mainly measures bug fixing, not full software development; BrowseComp and similar datasets and frameworks were overly specific. A model can score highly on a benchmark yet fail to understand the framework when connected to Claude Code or OpenClaw, leaving it short of industrial usability.
• The human-model division of labor has also changed in the real Agent era. People no longer point out code errors line by line; they add constraints, clarify requirements, participate in architecture design and provide enterprise business logic. The model handles a more complete action chain.
• MiMo’s latest optimization therefore “completely abandoned” old benchmarks for a time. At the initial paradigm break, a huge quality gap was obvious by feel; deeper into the new regime, the team still needs fine-grained evaluation.

18. Skills Give Models Organizational Intelligence That Does Not Exist on the Internet

• The value of Skills is not another block of generic knowledge, but execution standards for complex processes that improve accuracy on high-workflow-complexity tasks. Many such standards come from internal enterprise collaboration and do not exist in internet pretraining data at all.
• People can teach an Agent these implicit rules by completing several tasks over multiple rounds, after which the Agent can write them up as skills. 罗福莉 sees this class of knowledge as an important supplement to pretraining.
• Skills first appeared in Claude, but OpenClaw made them popular and encouraged more advanced users to contribute data inaccessible to models. They provide a natural-language interface for human-machine co-creation and one mechanism through which collective intelligence can scale.

19. The OpenClaw Moment Is a Longer, More Chaotic Diffusion Curve

• Unlike o1 or R1, which had a clear capability threshold, OpenClaw has no single moment when a score suddenly crosses a line. Its preceding evolution lasted a long time, while its subsequent capabilities are gradually spreading into more people’s real work.
• 罗福莉 believes its impact “will last longer and go deeper,” but energy takes time to diffuse. Many tasks lack the clear ground truth of mathematics or code; value accumulates slowly and then jumps at a threshold.
• Claude Opus 4.6 was a visible step change, but Anthropic had already been moving in that direction for at least 2 years. OpenClaw open-sourced black-box practice, allowing more people to see, modify and accelerate the framework for the first time.

20. Open Source Is Critical to Local Privacy and Hybrid Edge-Cloud Inference

• Asked whether ordinary people would hand large amounts of private information to Agents, 罗福莉’s answer is not to make the cloud framework absolutely secure, but to move simple, privacy-sensitive tasks to local inference.
• She envisions people carrying a chip or deploying one at home, with data remaining local. Only non-private, difficult, highly creative and highly complex tasks would go to the cloud, decoupling edge and cloud.
• The performance of 3B small models under complex frameworks convinces her that many local tasks require much smaller models than expected. But models, hardware and Agent architectures must evolve together; “no single company can do this well,” making open source even more important.

21. The Next Stage Is Not Another “OpenClaw-Like” Product but a Self-Evolving Framework

• 罗福莉 has tried roughly half of the comparable Chinese products and finds them “more or less the same”: they mainly turn OpenClaw into a similar product form, allowing users to access different models and framework types, without visibly outrunning the open-source community’s iteration speed.
• The real gap has 2 layers: first, self-iteration by the Agent or framework; second, mutual evolution between Agents and people. MiMo is currently more focused on training better Agent models and making Agents adapt to those models.
• She does not see a single core defect in OpenClaw that can be fixed in one stroke. “It’s simply that, because of time, this is where we’ve reached.” Models, frameworks, human feedback and evaluation will all accelerate together.
• The program discussed OpenClaw’s later sale to OpenAI. 罗福莉 believes models and frameworks do need deep coupling, but the good news is that the open-source character remains and the “genetic spark of collective evolution” has not disappeared.

22. The Main Theme of 2026 Is Continued Breakthroughs in High-Value Productivity

• 罗福莉 defines the current period as “an era of accelerating productivity transformation.” The most immediate change is that people no longer need to do many tasks themselves, forcing them to rethink their own meaning and value.
• If the goal is to improve frontier-model capability, having Agents replace higher-value tasks matters more. That means longer context, more tokens and more complex resource scheduling, and may ultimately replace the intelligence of the most capable people.
• If the goal is to benefit society broadly, the route is different: more general tasks, stronger multimodality, finer environmental understanding and lower cost matter more than serving a small set of high-value use cases.
• Both routes are constrained by economics. “You can’t spend $1,000 to complete one task.” End-to-end completion rate, speed and replacement cost will determine whether intelligence becomes a genuine productivity revolution.

23. Small Models Will Improve, but They Are Not the Main Story of 2026

• 罗福莉 sees the continued shrinking of models at the same capability level as part of evolution. A model that currently activates 3B parameters could reach Claude Opus 4.6-level performance in a year, while a million-token context could cost only RMB1-2.
• Smaller models are faster and cheaper, and become highly attractive when paired with modifiable Agent frameworks. She considers an edge-model boom a certain trend, but calls it a continuing “secondary line.”
• The core of 2026 is high-productivity use cases, long-horizon tasks and more complex collaboration—not simply compressing every model onto the edge.

24. Multi-Agent Currently Improves Cost and Speed, Not the Intelligence Ceiling

• High-complexity tasks are difficult for a single Agent to complete. Multiple Agents can divide work in parallel, saving time and cost, but 罗福莉 has not yet seen stable evidence that they raise the upper bound of final task completion.
• Her assessment of much of the multi-agent work in the market is that “some of it is fake.” The basic form already exists and she uses it comfortably, but it is delivering mainly on efficiency rather than producing outcomes beyond a single Agent’s reach.
• Each Agent still needs a relatively independent context, along with the ability to self-evolve and communicate with the others. She remains cautious on whether collaborative RL can break the ceiling: “Maybe not. Maybe.”
• Startups could nevertheless shrink dramatically as a result. She expects that within this year, one person supporting multiple digital employees will move from gimmick toward reality, provided model cost remains well below the value created by a real employee.

25. MiMo’s Goal Is Not DAU but the Point at Which Models Surpass Their Trainers

• 张小珺 noted that Chinese teams are diverging: Doubao, Yuanbao and Alibaba look more like internet products competing for DAU, while the Kimi team believes it has begun playing a different game.
• 罗福莉 says MiMo cares more about “when we will be able to surpass ourselves.” DAU does not directly produce models capable of replacing researchers; token consumption, completion of high-value tasks and complex context are closer to that goal.
• Her imagined model can access data, use a GPU cluster, define its own evaluation standards, train models and validate the results. Team-management capability would also mean mobilizing the intelligence of other Agents.
• That reopens the boundary of “the model as product.” Agent frameworks give models greater product power, while traditional UIs and peripheral functions become simpler; model companies will find it increasingly difficult to claim they do not make products.

26. AI Training AI Could Move from Impossible to Reality Within 1-2 Years

• 罗福莉 once believed that training models required the long contexts accumulated by researchers over many years, making that knowledge difficult to transmit to AI. Recently, however, she found that giving a model the relevant recent context can allow it to reconstruct a researcher’s own development path.
• Once the model has access to a similar context, she feels it can be “as intelligent as you,” and may even train a model as capable as the team can train. The uncomfortable implication is that research, once considered the hardest work to turn into skills, is beginning to look replicable.
• The next step is for models to train stronger models, improving by “stepping on their own left foot with their right.” Her view is that AI will first absorb everyone’s intelligence and then create new intelligence itself, and that this will definitely happen within 1-2 years.
• She acknowledges that vendors will not launch “having AI train AI” as a standalone project. It is a higher-order endpoint, but all frontier models will ultimately converge toward it.

27. MiMo-V2’s “Three-Model Ambush” Came from an Internal Awakening

• MiMo-V2-Pro, Omni and TTS were described as a “silent ambush” because the team itself had not expected capabilities to improve so quickly. The shift was invisible externally but happened rapidly inside the company after the OpenClaw shock.
• A year earlier, the multimodal and speech efforts were still isolated models: one understood video, another generated speech, but neither knew how to organize the pieces. OpenClaw gave 罗福莉 a clear first view of where each model should operate in a task chain.
• The team immediately redesigned post-training across all directions around the Agent paradigm. She believes that placing several MiMo models inside one framework will gradually create a combined advantage in performance, cost and speed.

28. Pro, Omni and TTS Handle Cognition, Perception and Expression

• 罗福莉’s division of labor is straightforward: Pro handles understanding, cognition and complex orchestration; Omni handles perception; TTS handles voice output and expression. The goal is to replace more parts of human life and work.
• They were not forcibly combined into one giant unified model, mainly for cost and speed reasons. Speech generation does not need to call a 1T model, while whether multimodal understanding merits more parameters must be judged task by task.
• Training within the same ecosystem has a hidden benefit: the main framework knows these models’ backgrounds and knowledge bases and can delegate more confidently without worrying that their foundations are completely misaligned.
• The performance gap between this internal combination and assembling models from other vendors is currently “very slight.” 罗福莉 expects it to widen because the real competition is the combined result across performance, cost and efficiency.

29. Flash and Pro Were Designed for Non-Chat Efficiency from the Start

• MiMo-V2-Flash and Pro were trained concurrently and have similar structures. The team’s clearest design bet from day one was that non-Chat modeling quality and inference efficiency had to be strong at the same time.
• The team vaguely anticipated that long context would create intelligence, but did not foresee a product form like OpenClaw. A low-cost, high-speed architecture is what could eventually scale 1T to 10T or even 100T without inference cost becoming prohibitive.
• 罗福莉 opposes hard-coding too many objectives into an architecture. After 6 or 12 months of post-training, the early assumptions about scenarios, chips and context length may all be obsolete. “Too many constraints” make the structure less adaptable.
• Flash was used to validate that the full inference data pipeline was sound, not to train Flash first and then decide what Pro should be. The 2 advanced in parallel; Pro, at the 1T scale, encountered more numerical-stability problems and finished later.

30. Hybrid Attention Leaves Structural Headroom for Future Workloads

• MLA, which was more mainstream at the time, was an elegant choice for the Chat era: it reduced KV cache and pushed memory access and compute toward a suitable balance for specific hardware.
• 罗福莉’s objection is that MLA already sits close to the boundary between compute-bound and memory-bound operation. Adding MTP could push it further into compute-bound territory, making it difficult to use spare compute for acceleration.
• Hybrid attention lets the team reduce KV cache with a sliding window and use MTP to fill idle compute. The ratio between full attention and sliding-window attention can also be adjusted later.
• The 2 designs represent different strategies: one locks in chips, context and inference-parallelism methods early to optimize a specific scenario; the other leaves room for long post-training cycles and an unknown Agent paradigm.

31. MTP Converts Idle Compute into Speed Without Adding Hallucinations

• MiMo did not choose MTP because of an initial ideological commitment. Midway through training, the team was designing a parallelism scheme for its own inference cards and discovered that spare compute was far greater than expected; MTP was a natural way to use it.
• Flash can reach roughly 100–150 TPS in inference, while Pro can reach roughly 60–100 TPS depending on cost. 罗福莉 says users’ most immediate impression of MiMo is that it is “fast,” and after experiencing a model with comparable intelligence at higher speed, it is hard to go back.
• MTP predictions are verified and accepted only when they hit, so predicting more tokens does not itself create additional hallucinations. With a high hit rate, it emits more tokens in less time, raises GPU utilization and lowers cost per token.
• The training strategy adds one MTP layer during pretraining to improve the base model, then trains more layers in mid-training and post-training for greater acceleration at inference.

32. Pro Pushes Full Attention and Sliding Windows to a 7:1 Ratio

• The team initially used a ratio of roughly 5:1, then pushed Pro to a 7:1 hybrid ratio. The experimental conclusion was that the absolute number of full-attention layers matters greatly, but the ratio can become sparser as the model grows.
• Larger models generally tolerate attention sparsity better, while excessive sparsity creates a clear drop in smaller models. 罗福莉 stresses that this is an experimental result from MiMo, not a fixed standard.
• The design goal is to avoid scaling full-attention layers in proportion to Pro’s significantly larger parameter count, keeping long-context efficiency close to Flash while raising the intelligence ceiling.
• Lower KV cache also allows more levels of caching. Once an Agent’s long-horizon interaction hits the cache, inference cost falls. This is where the architecture happens to fit the Agent workload particularly well.

33. A 1M Window Can Be Trained; Finding Enough Real 1M Data Is Harder

• 罗福莉 believes that if there were 1T tokens of high-quality 1M-context data, the model could learn 1M capability as long as loss continued to decline.
• The real problem is scarcity of scenarios and the extremely high cost of constructing them. It is difficult to obtain 1M data with sufficiently dense supervision, so MiMo’s long-context performance is improving incrementally rather than by simply inflating the advertised window.
• She says all model architectures could theoretically be trained to handle 100T context, but vendors do not offer inference that long not only because of quality concerns, but because the cost is high enough that users “don’t want to use it.”

34. Flash’s Cost-Plus Pricing Logic No Longer Fits the Agent Era

• At launch, Flash pursued extreme value, with API pricing of about $1.01 per 1M input tokens and $0.30 per 1M output tokens, while offering unusually high speed for the time.
• 罗福莉 says there was no single trick behind that result: the model architecture, inference architecture, MTP acceleration and infrastructure all had to be solid. “Basically, we did everything that needed to be done.”
• Pro, however, abandoned pricing based solely on inference cost. In the Agent era, value also comes from post-training, context understanding and completion rates on complex tasks; pricing should reflect the value ultimately created, not merely the underlying GPU cost.

35. 1T Is the Ticket to Getting Close to Opus 4.6, Not Victory Itself

• 罗福莉 previously trained DeepSeek-V3 at roughly 600B-700B parameters, so the next step naturally would not be another model in the same range. Given the compute available at the time, 1T was the upper end of what was reachable.
• Her empirical judgment is that bringing Agent capability close to Claude Opus 4.6 requires at least 1T total parameters. She does not know Opus’s own size and explicitly limits the conclusion to her practical experience.
• With total parameters held constant, a larger active parameter count raises inference cost; capability and economics remain a trade-off. 1T provides potential, while agile post-training and an adapted framework determine whether that potential is realized.
• She uses that framework to assess the domestic landscape: several vendors already have 1T-plus base models, and some teams have demonstrated post-training agility, but at the time few had both cards in hand.

36. Compute Allocation in the Agent Era Has Shifted from 5:1 to 1:1

• In the Chat era, pretraining-to-post-training compute might have been 3:1 or even 5:1. In the Agent era, longer post-training cycles likely move frontier teams toward 1:1.
• 罗福莉’s fuller allocation is roughly 3:1:1 across research, pretraining and post-training. Formal pretraining and post-training consume similar compute, while the additional research GPUs must at least exceed the total used for formal training.
• A single Flash or Pro training run requires several thousand GPUs, but research resources for architectural exploration, mid- and late-stage algorithms and parallel experiments may be 3-5x the formal training requirement.
• Agents also make ideas and code appear too quickly, shifting the bottleneck from people to GPUs. Inference-card demand is even larger and must be calculated separately based on user numbers and token consumption in high-value scenarios.

37. The Hardest Part of Training a 1T Model Is Treating Anomalies as Problems

• Large-model training commonly produces loss spikes, exploding activations and extreme expert-load imbalance. 罗福莉 refuses to treat a spike as normal noise, because an anomalous update at one step can directly “kill” parameters or an expert.
• The team monitors inputs and outputs at every layer, activations and expert loads, then investigates layer by layer through structural ratios, communication operators, normalization and hardware failures. At the end, they may even wonder whether the cause is sunspots.
• Temporary measures can include clipping anomalous values, adding normalization or borrowing QK clipping. But normalization can damage performance, so the team must determine whether a fix is a permanent structural change or merely a bridge into a stable regime.
• Training has been halted for days and at times 2 weeks because of these issues. Stopping several thousand GPUs for a day can cost RMB1M-2M or RMB2M-3M, but the team still chooses to solve the problem. She says there is no firm release date, yet she dreams about “why loss spiked again.”

38. Small Teams Are Better Than Layered Coordination at Debugging Ambiguous Training Failures

• The number of people truly involved across training, inference and data is not large. What matters is whether several algorithm, training-framework and inference specialists can debug together, not the overall headcount.
• Cross-team investigations at large companies can become extremely inefficient. In a small team, people bring different hypotheses and solve the problem together directly. 罗福莉 believes large teams may be poorly suited to tracking down “things that might be problems.”
• MiMo has no rigid release target: “We’ll release when the model is trained.” That flexibility lets the team stop an expensive run in exchange for a more stable and understandable training process.
• She sees this hands-on, collective problem-solving approach as culture rather than management: “Everyone solving the problem together is enough.”

39. MiMo Uses a Flat 100-Person Organization to Run Multiple Model Pipelines

• The team has about 100 people covering data collection, quality, pretraining infrastructure, post-training, development, product, labeling, language, multimodality and speech; interns also make up a significant share.
• Only roughly 20-30 people, at most 30-40, may be directly iterating on a given model generation. Training itself is relatively small-scale, data requires only several people, and an experienced infrastructure team is then needed to stabilize the cluster.
• The team is not divided into fixed pretraining and post-training groups. Projects can have a driver, but that person does not exercise absolute control over participants; Xiaomi has job grades, but the internal organization is completely flat.

40. Egalitarianism Is Not a Cultural Slogan but a Way to Avoid Suppressing Model Innovation

• 罗福莉 believes fixed divisions kill creativity across stages. Many pretraining data specialists naturally move into post-training because data intuition, working backward from results to algorithms and an emphasis on diversity are all transferable.
• Traditional post-training specialists tend to focus on individual scenarios, while pretraining specialists naturally focus on diversity. The 2 perspectives complement each other better than building a new team from rigid job profiles.
• She does not believe that greater seniority necessarily means intelligence superior to everyone else’s. “That definition is very strange.” The most important leaders should not feel that “without me, everything falls apart.”
• A project can have a driver, but organizational authority cannot be equated with judgment. Her management principle is to “let passion drive the work,” allowing people to move organically around goals they believe in.

41. Environment Matters More Than Experience; Capabilities Can Be Learned in 1-4 Months

• Many team members had never trained large models before joining, and only roughly one-quarter to one-third had worked with 7B or 14B models. She does not believe that experience transfers directly to training at 1T.
• 罗福莉’s judgment is clear: “These capabilities can genuinely be learned quickly—in at most 1-2 months, or 3-4 months if it takes longer.” The condition is an environment with high standards and strong peers working together on real problems.
• She cares more about a candidate’s initial ceiling than their current point after being “supervised” by past experience. Technical fundamentals, curiosity, passion and the ability to turn an idea into reality matter more than labels.
• PhDs and current PhD students make up roughly 55% of the team, but she sees that number as rigid. Recently she has leaned toward hiring second- and third-year undergraduates, who are less contaminated by old paradigms and more willing to hand ideas to an Agent for validation.

42. She Has Moved from Confidence to Caution on Whether Multimodality Creates Higher Intelligence

• MiMo-V2-Omni natively supports joint understanding of video, audio, images and text and aims to bring Agent capability close to that of a language model. Its parameter count is below Pro’s, but the exact size has not been disclosed.
• Two months ago, 罗福莉 strongly believed that multimodality itself would produce greater intelligence; during training, she began to question that view. Omni may remain “completely unchanged” on benchmarks, providing no proof that expanding perceptual dimensions necessarily improves reasoning.
• Real-world use still produces 2 signals: because of its video training, Omni has more world knowledge, while its sensitivity to subtle environments, emotional intelligence and insight may also exceed those of a larger pure language model.
• She is not prepared to call multimodality a necessary route to AGI, especially since Agents can already orchestrate multiple specialist models. Her newer hypothesis is that perception alone may not increase intelligence, while unifying understanding with dynamic generation might.

43. TTS Is Betting on Generalization Through Discrete Tokens, Not Short-Term Architectural Convenience

• MiMo’s audio team is discretizing continuous waveforms into text-like token IDs and using multiple layers of RVQ to construct high-dimensional representations, with the goal of reconstructing the signal as losslessly as possible.
• The approach emerged later than continuous features and requires a larger pretraining investment. Part of the motivation is “the obsession of people who build LLMs”: discrete, simple and easy-to-define supervision.
• TTS is SFT- and RL-trained on only a few stereotyped styles such as fast, slow, happy and sad, yet can follow complex natural-language descriptions of style and infer emotion and prosody from text meaning. 罗福莉 sees this as the generalization enabled by a simple architecture plus massive-scale training.
• She also acknowledges that the model’s “ceiling is extremely high, while the floor still needs work.” It remains occasionally unstable; the current time-limited free API is mainly for experimentation and is not guaranteed to be production-ready, though she expects it to reach production quality soon.

44. She Has Pulled the AGI Timeline Forward to Within 2 Years but Refuses to Treat the Future as Fixed

• 罗福莉 estimates that the journey toward AGI is about 20% complete today, could reach 60%–70% this year and may overturn the way most people work within 2 years. Changes to daily life will be slower because that requires robots to move from screens into the physical world, while hardware, batteries and dexterous hands are advancing more slowly.
• Chinese teams with 1T-plus base models could, if technically agile, catch up to today’s Claude Opus 4.6 in 2-3 months rather than to the Claude of 2-3 months from now. Agent frameworks will improve rapidly in parallel, inference demand could grow by several times to 10x, and storage and chip capacity will become binding constraints.
• She reviews the 3-year progression: ChatGPT in 2022 used conversation to unlock pretrained intelligence; open-source teams then used Llama to chase frontier closed models, Qwen emphasized scale and ecosystem, and DeepSeek advanced MoE, MLA and scaling under constrained resources. o1 and R1 pushed competition toward reasoning, while 2025 became the year the Chat and Agent paths diverged.
• She repeatedly warns that her judgment may change: “Every day I may be disproving yesterday’s self.” Her guiding principle has shifted from the quant-era belief that “there is always a way to model prices” toward doing work aligned with her values and capable of replacing drudgery. She has considered building a public-interest organization to support Chinese basic research and let people and AI continue exploring together.